Retire the networking bundle and what the control plane stops shipping (hq ADR 0226)

networking required mesh-wireguard and nothing else; machines are assigned the network directly.
module forget refuses a provided module, so a retired one is removed at start once no machine has it.
Guard route-proxy's public account directory against a reissue.

.gitignore

@@ -2,11 +2,4 @@
 
 # A built binary. The lab writes one here when pointed at the repository root by mistake;
 # built artifacts belong in build/, which is already ignored.
-/mesh-controller
-/postgres-provisioner
-/redis-provisioner
-# And where `go build` in cmd/mesh-controller leaves it (novox/hq issue 233).
-/cmd/mesh-controller/mesh-controller
-
-# No repository commits its dependencies (issue 1305).
-vendor/
+/mesh-builder

Dockerfile

@@ -0,0 +1,41 @@
+# The Go it builds with, pinned here because genesis builds this file with no arguments (novox/hq
+# issue 223) — the Makefile passes the same digest. A tag older than go.mod asks for is how
+# `make image` broke once before (issue 146).
+ARG GO_BASE=golang@sha256:8ac98ca534ac3f51e1f420a1dd2c15e74c75cfa0f23f3ad27eb5d7236c349a0c
+# The control plane's image — for genesis and the lab only. The mesh runs the controller as a Go
+# bundle the host starts as a process (module.json; novox/hq issue 213), and builds no image of it.
+# Genesis builds this file and raises it as the container the process replaces on the first push
+# (mesh-host internal/bootstrap, novox/hq issue 223).
+#
+# novox/hq ADR 0006: this image is pinned by digest in the bundle the host carries, fetched on a
+# machine where no mesh exists yet, and run before there is anything to check it against. So it
+# holds the program and nothing else — no shell, no package manager, no libc, nothing with a CVE
+# feed of its own. What a person has to audit before trusting a first node is one binary.
+#
+# There are no CA certificates in here on purpose. Nothing it does today makes an outbound TLS
+# connection to a public name: it reaches PostgreSQL on the machine it was raised on, and the
+# broker is verified against a fingerprint pinned in a token rather than against a public root
+# (novox/hq ADR 0004). Adding them "just in case" would put a trust store in the one image whose
+# whole argument is that it contains nothing to reason about.
+
+FROM ${GO_BASE} AS build
+WORKDIR /src
+
+# Dependencies first, so a change to the source does not refetch them.
+COPY go.mod go.sum ./
+RUN go mod download
+
+COPY . .
+ARG VERSION=development
+RUN CGO_ENABLED=0 go build -trimpath \
+        -ldflags "-s -w -X main.version=${VERSION}" \
+        -o /mesh-controller ./cmd/mesh-controller
+
+FROM scratch
+COPY --from=build /mesh-controller /mesh-controller
+
+# Numeric because there is no /etc/passwd to look a name up in. Nothing here needs to be root:
+# it opens outbound connections and writes nothing to its own filesystem.
+USER 65534:65534
+
+ENTRYPOINT ["/mesh-controller"]

Makefile

@@ -0,0 +1,141 @@
+# novox/hq ADR 0006 — the control plane, in Go.
+#
+# The image the bundle pins holds the program and nothing else, so the build is static and the
+# container is built FROM scratch. That is not a size optimisation: this image is fetched by
+# digest and run on a machine where no mesh exists to check anything, and everything in it is
+# something a person would have to audit.
+
+VERSION ?= $(shell git describe --tags --always --dirty 2>/dev/null || echo development)
+LDFLAGS := -s -w -X main.version=$(VERSION)
+
+# Where `make check` raises PostgreSQL. A high port and a throwaway container: nothing here
+# touches a database anybody else is using. Override PG_PORT if this one is taken -- the first
+# port chosen was already serving something that had been up for six days.
+PG_PORT      ?= 55532
+PG_CONTAINER ?= mesh-controller-check
+PG_IMAGE     ?= postgres:17-alpine
+export MESH_TEST_POSTGRES ?= postgres://postgres:check@127.0.0.1:$(PG_PORT)/postgres?sslmode=disable
+
+.PHONY: build image check test vet fmt postgres postgres-stop clean
+
+build:
+	CGO_ENABLED=0 go build -trimpath -ldflags '$(LDFLAGS)' -o build/mesh-controller ./cmd/mesh-controller
+
+# Tagged 'development' as well as by version, because the lab places images by name and a
+# scenario naming a version would have to be edited on every build. The version tag is what a
+# real bundle pins.
+IMAGE ?= mesh-controller:$(VERSION)
+DEV_TAG ?= mesh-controller:development
+
+# The Go base the image is built on.
+#
+# **`make image` was broken and stayed broken**, because the Dockerfile's fallback base was a Go
+# older than go.mod asks for: every build died at `go mod download` with "go.mod requires go >=
+# 1.26.0", and the pipeline never saw it because the pipeline passes the declared base in. Anybody
+# building the image by hand hit it and had to find the digest themselves (novox/hq 04-ISSUES/146,
+# what it cost).
+#
+# **Pinned here since the manifest stopped building an image** (novox/hq issue 213): the mesh builds
+# the controller as a Go bundle with its own toolchain, and only `make image` — genesis and the lab —
+# still needs a Go base. The digest is the one the manifest declared until then.
+GO_BASE ?= golang@sha256:8ac98ca534ac3f51e1f420a1dd2c15e74c75cfa0f23f3ad27eb5d7236c349a0c
+
+image:
+	@test -n "$(GO_BASE)" || { echo "no GO_BASE; pass GO_BASE=<image>"; exit 1; }
+	docker build --build-arg GO_BASE=$(GO_BASE) --build-arg VERSION=$(VERSION) -t $(IMAGE) -t $(DEV_TAG) .
+	@echo
+	@docker image inspect $(IMAGE) --format 'built {{.RepoTags}}  {{.Size}} bytes'
+
+# The builder ships as an image too, because it is a module the mesh assigns rather than a program
+# somebody starts on a machine by hand.
+BUILDER_IMAGE ?= mesh-builder:$(VERSION)
+BUILDER_DEV_TAG ?= mesh-builder:development
+
+builder-image:
+	@test -n "$(GO_BASE)" || { echo "no GO_BASE; pass GO_BASE=<image>"; exit 1; }
+	docker build --build-arg GO_BASE=$(GO_BASE) -f cmd/mesh-builder/Dockerfile -t $(BUILDER_IMAGE) -t $(BUILDER_DEV_TAG) .
+	@echo
+	@docker image inspect $(BUILDER_IMAGE) --format 'built {{.RepoTags}}  {{.Size}} bytes'
+
+# The provisioner ships as an image too, because it is the thing that makes a sealed credential
+# true on a machine -- and the mesh cannot, having discarded the plaintext.
+PROVISIONER_IMAGE ?= mesh-provision-postgres:$(VERSION)
+PROVISIONER_DEV_TAG ?= mesh-provision-postgres:development
+
+provisioner-image:
+	docker build --build-arg GO_BASE=$(GO_BASE) -f examples/postgres-provisioner/Dockerfile \
+		-t $(PROVISIONER_IMAGE) -t $(PROVISIONER_DEV_TAG) .
+	@echo
+	@docker image inspect $(PROVISIONER_IMAGE) --format 'built {{.RepoTags}}  {{.Size}} bytes'
+
+# The object store's provisioner, for the same reason: a bucket and a policy are not files, and
+# the mesh cannot make them -- it discarded the credential it would have to use.
+OBJECTSTORE_IMAGE ?= mesh-provision-objectstore:$(VERSION)
+OBJECTSTORE_DEV_TAG ?= mesh-provision-objectstore:development
+
+objectstore-image:
+	docker build --build-arg GO_BASE=$(GO_BASE) -f examples/objectstore-provisioner/Dockerfile \
+		-t $(OBJECTSTORE_IMAGE) -t $(OBJECTSTORE_DEV_TAG) .
+	@echo
+	@docker image inspect $(OBJECTSTORE_IMAGE) --format 'built {{.RepoTags}}  {{.Size}} bytes'
+
+# The cache's provisioner, for the same reason as the database's: an ACL user is not a file,
+# and the mesh cannot make one -- it discarded the credential it would have to use.
+REDIS_PROVISIONER_IMAGE ?= mesh-provision-redis:$(VERSION)
+REDIS_PROVISIONER_DEV_TAG ?= mesh-provision-redis:development
+
+redis-provisioner-image:
+	docker build --build-arg GO_BASE=$(GO_BASE) -f examples/redis-provisioner/Dockerfile \
+		-t $(REDIS_PROVISIONER_IMAGE) -t $(REDIS_PROVISIONER_DEV_TAG) .
+	@echo
+	@docker image inspect $(REDIS_PROVISIONER_IMAGE) --format 'built {{.RepoTags}}  {{.Size}} bytes'
+
+# The proxy that turns a route grant into traffic reaching a workload.
+PROXY_IMAGE ?= mesh-route-proxy:$(VERSION)
+PROXY_DEV_TAG ?= mesh-route-proxy:development
+
+proxy-image:
+	docker build --build-arg GO_BASE=$(GO_BASE) -f examples/route-proxy/Dockerfile -t $(PROXY_IMAGE) -t $(PROXY_DEV_TAG) .
+	@echo
+	@docker image inspect $(PROXY_IMAGE) --format 'built {{.RepoTags}}  {{.Size}} bytes'
+
+# The whole gate. Raises a database, runs everything against it, and takes it down again --
+# including when the tests fail, which is why the teardown is not conditional.
+#
+# **One package at a time (-p 1), and it is not about speed.** The live tests reach one bus, and on
+# it they assert, read and remove the mesh's own objects -- streams and consumers with fixed names,
+# because those names are the mesh's and a test cannot choose others. Two packages doing that at once
+# is one deleting a consumer the other is reading through, and the failure lands in whichever test
+# was reading, as "no response from stream". That reads as a bug in the code under test.
+check: fmt vet postgres
+	@go test -p 1 ./... ; status=$$? ; $(MAKE) postgres-stop ; exit $$status
+
+# Without a database the live tests skip rather than fail, so this is the honest subset and not
+# the gate. Serialised for the same reason check is: a bus may be configured even when a store is not.
+test:
+	go test -p 1 ./...
+
+vet:
+	go vet ./...
+
+fmt:
+	@unformatted=$$(gofmt -l . 2>/dev/null) ; \
+	if [ -n "$$unformatted" ] ; then echo "not gofmt'd:" ; echo "$$unformatted" ; exit 1 ; fi
+
+postgres:
+	@docker rm -f $(PG_CONTAINER) >/dev/null 2>&1 || true
+	@docker run -d --name $(PG_CONTAINER) -e POSTGRES_PASSWORD=check \
+		-p 127.0.0.1:$(PG_PORT):5432 $(PG_IMAGE) >/dev/null
+	@printf 'waiting for postgres'
+	@for i in $$(seq 1 60) ; do \
+		if docker exec $(PG_CONTAINER) pg_isready -U postgres >/dev/null 2>&1 ; then \
+			echo ' — ready' ; exit 0 ; fi ; \
+		printf '.' ; sleep 1 ; \
+	done ; \
+	echo ' — never came up' ; docker logs $(PG_CONTAINER) | tail -20 ; exit 1
+
+postgres-stop:
+	@docker rm -f $(PG_CONTAINER) >/dev/null 2>&1 || true
+
+clean:
+	rm -rf build/

README.md

@@ -1,7 +1,218 @@
 # mesh-controller
 
-The controller ended under issue 1126. Its jobs moved to the parts that ADR 0590 and ADR 0593 name: the indexer, the
-assigner, the provisioner, the grants part and the node-engine. Its last module, route-proxy, moved to mesh-networking.
+**Tier 2 of Novox Mesh — the control plane.** Everything that needs to know about more than one
+node.
 
-This repository holds no code and is archived once the delivery of the group that emptied it (release sdk1) is done.
-Its history stays here for reading.
+That is the whole test, and it draws the line the host cannot: the host applies and does not
+decide, *because deciding needs knowledge one machine does not have*. Which nodes should run the
+store, which peers belong in an overlay, whether a node has been unreachable for a week — nobody
+on a single machine can answer any of them.
+
+The reasoning lives in [novox/hq](https://git.novox.be/novox/hq); this repository carries no
+argument that is not settled there.
+
+## What it is not
+
+- **Not the thing that changes machines.** It decides; the host applies. It never reaches into a
+  node except through the host, over the link, in a bounded vocabulary.
+- **Not a database.** There is no mesh database. Each context owns its own store and nothing
+  outside a context touches it — including nodes, which hold no credential to any of them.
+- **Not privileged.** It has no more access to a machine than a declaration can express.
+
+## What exists today
+
+**One context of seven, and one of the things it will do.**
+
+| | |
+|---|---|
+| `inventory` | node records and enrolment tokens — **built, as far as identity** |
+| `identity` | the control plane's signing key, and the keys nodes are known by — **built** |
+| `config`, `connectivity`, `provisioning`, `delivery`, `observability` | not built |
+| the interface every surface speaks to | not built; its shape is not decided |
+
+```
+mesh-controller migrate                    bring each context's schema up to date
+mesh-controller node add <name>            create a node record
+mesh-controller node list                  the nodes this mesh knows about
+mesh-controller token issue --node <name>  a one-time right to join, for an existing record
+mesh-controller token issue --new <name>   create the record and issue for it
+mesh-controller identity show              this control plane's signing key
+mesh-controller broker show                where the broker is, and what to expect there
+mesh-controller version                    what this binary is
+```
+
+`migrate` is **step 3 of the foundation bootstrap** — the step the first node cannot get past, run
+against a database raised moments earlier from the bundle the host carries.
+
+### Tokens, and what they are missing
+
+A token is **a one-time right to join, issued for a node record** — which is where re-enrolment is
+decided, since what an identity binds to is settled when the token is made rather than when it is
+presented.
+
+What is built: the secret is 256 bits from the system's random source, shown once, and **stored
+only as a hash**, so a copy of this database is not a set of working credentials. It is usable
+exactly once and only before it expires, and both are read from the row rather than from a status
+something would have had to write. Issuing again for the same node invalidates the outstanding
+one — two live tokens are two machines able to join as the same node.
+
+Redemption is a single statement that both finds a live token and spends it, so eight concurrent
+attempts on one secret produce exactly one winner. There is a test that runs them.
+
+**A token now carries three of its four parts**, and is one line of base64 a person can copy. The
+signing key is real: an Ed25519 key this control plane generates once and keeps, whose public half
+travels in every token. A node believes a declaration because it carries a signature that key made
+— and pinning only the broker would not do, because it would make the control plane's authority
+transitive, so a compromised broker could forge declarations, and since the host applies whatever
+the link delivers that is the whole machine.
+
+**All four parts are built.** Given `MESH_BROKER_ADDRESS` and `MESH_BROKER_CERTIFICATE`, a token
+carries everything ADR 0004 requires. Without them it carries two, and `token issue` prints it
+**and names what is missing** rather than producing something that looks complete and cannot be
+used.
+
+**The fingerprint is derived from the certificate, never configured.** A configured one can drift
+from the certificate it describes, and a drifted pin is worse than none: every node issued a token
+during the drift refuses to connect, and the failure looks like an attack. It is computed over the
+DER bytes — what a client actually sees on the wire — so the same certificate re-wrapped with
+different line endings still produces the same pin.
+
+### Two contexts, and the rule between them is real
+
+`identity` is the second context and it exists partly to test a claim this repository had made and
+never checked: that a context reaches only its own store. It holds `MESH_STORE_IDENTITY`;
+`inventory` holds `MESH_STORE_INVENTORY`; there is no setting that reaches both and no way to ask
+for one. Run `migrate` with only one and it stops, naming the grant it does not have.
+
+A token needs a node record from one and a signing key from the other. Neither reads the other's
+store — the process holding both grants asks each for its part.
+
+### How a node is known
+
+**The node generates a keypair; the mesh records the public half.** The same principle as SSH, and
+the same rule `08-connectivity` already applies to the overlay keys — which is where this was
+settled all along, though it took being asked directly to notice.
+
+Only the public half is ever stored, and that is the property worth having: **a copy of this
+database grants nothing.** It is a list of who to believe, not a set of credentials, which is what
+makes *compromise of a node is compromise of that node* literally true.
+
+Exactly one key is live per node. Re-enrolment revokes the one it replaced, in the same
+transaction — two live identities for one node record is the stolen-laptop case, with the replaced
+machine still believed. The database enforces it as well as the code, and there is a test running
+six concurrent enrolments that fails when the constraint is removed.
+
+**Not the machine's SSH host key**, though that was the obvious economy. Host keys are regenerated
+by reinstalls and image clones, which would silently un-enrol a node; their lifecycle belongs to
+sshd rather than the mesh; and a partial host has no SSH daemon at all.
+
+## Reaching a store
+
+**A context is granted only what it exclusively owns.** No shared writes, no read-only role on
+another context's store, and no connection string that reaches more than one.
+
+That is a rule about credentials, so it is built as one. There is no mesh-wide connection setting
+and no way to ask for one:
+
+```
+MESH_STORE_INVENTORY=postgres://…/inventory
+```
+
+A process granted `inventory` holds that variable and no other. Reaching another context's store
+is not a matter of restraint — it has no address for it and no credential to present. And it is
+how the rule is *checked*: what a context can reach is visible in the declaration that runs it,
+as the list of variables it was given.
+
+Each context's database is named after the context. There is deliberately no database named for
+the mesh as a whole.
+
+## Migrations
+
+Numbered, embedded in the binary, applied in order, each in a transaction with the row recording
+it. The applying is four lines; the rest is refusals, and the refusals are the point:
+
+| it stops when | because |
+|---|---|
+| a migration that ran has since been edited | the database holds the old version, the repository holds the new one, and nothing holds the difference |
+| a migration is numbered below one that already ran | usually two branches taking the same next number — applying it now runs the schema in an order nobody tested |
+| two migrations share a number | order is the entire guarantee, and two files with one number have none |
+| a file in the directory is not a valid migration name | a misnamed migration would otherwise never run and nothing would say so |
+| a migration is empty | it records that something happened and changes nothing, which cannot be told from a mistake |
+
+All of them stop rather than warn. At the moment any of them is true, nobody knows what the
+database contains, and there is no correct guess about a schema.
+
+Running it again does nothing. Two copies running at once take an advisory lock, so a restart
+during a slow migration does not become two runners racing.
+
+## Building
+
+```
+make build      the binary
+make image      the container image
+make check      gofmt, vet, and every test against a real PostgreSQL
+```
+
+`make check` raises a throwaway PostgreSQL in a container and takes it down afterwards, including
+when the tests fail. Without one the tests that need a database **skip and say so** rather than
+passing quietly — `make test` is the honest subset, not the gate.
+
+Tests run against a real database rather than a fake because what is being tested *is* the
+database's behaviour: that DDL is transactional, that an advisory lock serialises, that a
+checksum mismatch is caught against a record PostgreSQL actually kept. A fake would assert that
+the fake behaves as expected.
+
+Every test here has been confirmed to fail when the behaviour it defends is removed. Two did not,
+when first written, and both are now commented with what they were missing.
+
+## Checks that cross into the host's repository
+
+Two things are agreed between this repository and `novox/mesh-host`, and each is a separate struct
+on each side. A field renamed on one of them fails **silently** — the crossing succeeds and
+something is simply absent — so both are checked by handing one side's real output to the other's
+real parser. Neither runs by default; each skips with a reason, because a repository that fails
+without its neighbour checked out is a repository nobody can build.
+
+**What this mesh sends, read by the host that receives it:**
+
+```
+mesh-controller:  ./build/mesh-controller plan <node> --json > /tmp/d.json
+mesh-host:     MESH_EMITTED=/tmp/d.json go test ./internal/declaration/ -v
+```
+
+**What a node says when it joins, read by this mesh:**
+
+```
+mesh-host:     MESH_ENROL_OUT=/tmp/enrol.json go test ./internal/link/
+mesh-controller:  MESH_ENROL=/tmp/enrol.json make check
+```
+
+The second does more than compare shapes: it seals something to the key that arrived and opens it
+with the private half the host kept. A key that is present, correctly named and simply *wrong*
+passes every check that only looks at the message.
+
+## The image
+
+**The mesh no longer runs the controller from it** (novox/hq issue 213). The module declares a Go
+bundle, `controller`, which the host on the controller's machine unpacks and runs as the process
+`mesh-controller` under the account of the same name (ADR 0188 §1, §3). The image stays for what
+still runs a container of the controller: genesis, which raises the first controller from it and
+installs the module from its manifest (mesh-host `internal/bootstrap`), and the lab. Neither is the
+mesh's own build any more — `make image` builds it.
+
+`FROM scratch`, holding one statically linked binary and nothing else — no shell, no package
+manager, no libc, no CA certificates.
+
+Not a size optimisation. The bundle a host carries pins this image by digest, and it is fetched
+and run on a machine where no mesh exists to check anything and a person is expected to have read
+the bundle and believed it. Everything in the image is something that person would have to audit.
+
+## Where the reasoning lives
+
+| | |
+|---|---|
+| what the control plane is | novox/hq ADR 0006 |
+| what it takes to run one, and why Go | novox/hq ADR 0006 |
+| a context owns its store, exclusively | novox/hq ADR 0008 |
+| schema changes are numbered migrations | novox/hq ADR 0013 |
+| a test defends a decision | novox/hq ADR 0017 |

cmd/mesh-builder/Dockerfile

@@ -0,0 +1,22 @@
+ARG ALPINE_BASE=alpine:3
+ARG GO_BASE=golang:1.25-alpine
+# The builder, as a module ships one.
+#
+# Not FROM scratch, unlike the control plane: this one runs git and a container client, so it
+# needs a filesystem with them on it. That is the honest cost of a machine whose job is to build —
+# and it is why building is a MODULE on a machine that has a runtime rather than something the
+# control plane does (novox/hq ADR 0005).
+FROM ${GO_BASE} AS build
+WORKDIR /src
+COPY go.mod go.sum ./
+RUN go mod download
+COPY . .
+RUN CGO_ENABLED=0 go build -trimpath -ldflags '-s -w' -o /mesh-builder ./cmd/mesh-builder
+
+FROM ${ALPINE_BASE}
+# git to clone what it is asked to build, and the docker client to build and push it. The daemon
+# is the machine's, reached through its socket — a build machine shares the runtime it was given
+# rather than running one inside itself.
+RUN apk add --no-cache git docker-cli
+COPY --from=build /mesh-builder /usr/local/bin/mesh-builder
+ENTRYPOINT ["/usr/local/bin/mesh-builder"]

cmd/mesh-builder/holder.go

@@ -0,0 +1,386 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/nats-io/nats.go/micro"
+
+	"github.com/novox/mesh-controller/internal/builder"
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// What a holder of the build seat answers for, on its own machine (novox/hq ADR 0219).
+//
+// **The queue is the controller's; the build running here is this machine's.** The controller can
+// see and change what waits in the seat's queue, but an ask a holder already took is a process tree
+// on this machine and containers in this machine's runtime, and only this machine can end them. So
+// the holder serves four verbs on the seat's subjects for this machine: what it is building, kill
+// it, pause, resume.
+//
+// **One build at a time** (ADR 0190), which is also what builder.Said assumes — a package global
+// set per build — so "the build running here" is one or none, and kill names it by id so a call
+// that arrives as one build ends and the next begins cannot end the wrong one.
+
+// holder is this machine's state as a holder of the build seat.
+type holder struct {
+	on, seat string
+	// workspace is where the paused flag is kept, so a holder restarted while paused stays paused
+	// rather than silently taking work again.
+	workspace string
+	// say publishes this machine's state: whether it takes work (link.HolderState).
+	say func(link.HolderState) error
+	// remove runs what a kill needs outside the build: the containers left behind.
+	remove builder.Runner
+
+	mu      sync.Mutex
+	paused  bool
+	running *running
+}
+
+// running is the build this machine is doing.
+type running struct {
+	request link.BuildRequest
+	step    string
+	started time.Time
+	cancel  context.CancelFunc
+	killed  bool
+	// returned is the build's own work having ended, before a kill or not; outcome is what was then
+	// announced, and announced whether it went out.
+	returned  bool
+	outcome   string
+	announced bool
+	done      chan struct{}
+}
+
+// pausedFile is where the flag lives in the workspace.
+func pausedFile(workspace string) string { return filepath.Join(workspace, ".mesh-builder-paused") }
+
+// newHolder reads the paused flag the workspace keeps.
+func newHolder(on, seat, workspace string, say func(link.HolderState) error) *holder {
+	h := &holder{on: on, seat: seat, workspace: workspace, say: say, remove: plainRun}
+	if _, err := os.Stat(pausedFile(workspace)); err == nil {
+		h.paused = true
+	}
+	return h
+}
+
+// Paused is asked by the taking loop before every fetch.
+func (h *holder) Paused() bool {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	return h.paused
+}
+
+// setPaused records the flag in the workspace first and then in memory, so what this holder says
+// it is and what it would be after a restart never differ.
+func (h *holder) setPaused(paused bool) error {
+	path := pausedFile(h.workspace)
+	if paused {
+		if err := os.MkdirAll(h.workspace, 0o755); err != nil {
+			return err
+		}
+		if err := os.WriteFile(path, []byte(time.Now().UTC().Format(time.RFC3339)+"\n"), 0o644); err != nil {
+			return fmt.Errorf("cannot keep the paused flag in %s: %w", path, err)
+		}
+	} else if err := os.Remove(path); err != nil && !errors.Is(err, os.ErrNotExist) {
+		return fmt.Errorf("cannot remove the paused flag %s: %w", path, err)
+	}
+	h.mu.Lock()
+	h.paused = paused
+	h.mu.Unlock()
+	h.announce()
+	return nil
+}
+
+// announce says whether this machine takes work. Never fatal: the flag is kept either way, and the
+// controller reading an older state is a plan read as late rather than a build lost.
+func (h *holder) announce() {
+	if h.say == nil {
+		return
+	}
+	if err := h.say(link.HolderState{On: h.on, Paused: h.Paused(), At: time.Now().UTC().Format(time.RFC3339Nano)}); err != nil {
+		fmt.Fprintf(os.Stderr, "cannot say whether this machine takes builds: %v\n", err)
+	}
+}
+
+// stateWhilePaused says this machine's state at start, and again every while it stays paused, so
+// a pause outlives the events stream's retention.
+func (h *holder) stateWhilePaused(ctx context.Context, every time.Duration) {
+	h.announce()
+	tick := time.NewTicker(every)
+	defer tick.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-tick.C:
+			if h.Paused() {
+				h.announce()
+			}
+		}
+	}
+}
+
+// begin records the build this machine took, with the cancel that ends it.
+func (h *holder) begin(request link.BuildRequest, cancel context.CancelFunc) *running {
+	r := &running{request: request, started: time.Now(), cancel: cancel, done: make(chan struct{})}
+	h.mu.Lock()
+	h.running = r
+	h.mu.Unlock()
+	return r
+}
+
+// end clears it, and lets a kill waiting on it know it is over.
+func (h *holder) end(r *running) {
+	h.mu.Lock()
+	if h.running == r {
+		h.running = nil
+	}
+	h.mu.Unlock()
+	close(r.done)
+}
+
+// stepped records the step a build is at, for `current`.
+func (h *holder) stepped(r *running, step string) {
+	h.mu.Lock()
+	r.step = step
+	h.mu.Unlock()
+}
+
+// currentBuild is what `current` answers.
+type currentBuild struct {
+	On      string `json:"on"`
+	Paused  bool   `json:"paused"`
+	Running *struct {
+		ID         string `json:"id"`
+		Repository string `json:"repository"`
+		Path       string `json:"path,omitempty"`
+		Ref        string `json:"ref,omitempty"`
+		Step       string `json:"step,omitempty"`
+		Started    string `json:"started"`
+		Elapsed    string `json:"elapsed"`
+	} `json:"running,omitempty"`
+	Said string `json:"said"`
+}
+
+func (h *holder) current() currentBuild {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	out := currentBuild{On: h.on, Paused: h.paused}
+	taking := "taking builds"
+	if h.paused {
+		taking = "paused, taking no new build"
+	}
+	if h.running == nil {
+		out.Said = fmt.Sprintf("%s is building nothing; %s", h.on, taking)
+		return out
+	}
+	r := h.running
+	elapsed := time.Since(r.started).Round(time.Second)
+	out.Running = &struct {
+		ID         string `json:"id"`
+		Repository string `json:"repository"`
+		Path       string `json:"path,omitempty"`
+		Ref        string `json:"ref,omitempty"`
+		Step       string `json:"step,omitempty"`
+		Started    string `json:"started"`
+		Elapsed    string `json:"elapsed"`
+	}{r.request.ID, r.request.Repository, r.request.Path, r.request.Ref, r.step,
+		r.started.UTC().Format(time.RFC3339), elapsed.String()}
+	out.Said = fmt.Sprintf("%s is building %s (%s) at %s for %s; %s",
+		h.on, r.request.Repository, r.request.ID, orNothing(r.step), elapsed, taking)
+	return out
+}
+
+// The bounds a kill keeps: each pass removing containers, and the wait for the build to end between
+// them. The worst case — 15s, 20s, 15s — is inside what the controller waits for the answer
+// (killAnswer in the controller's queue.go, 75s).
+var (
+	killRemoves = 15 * time.Second
+	killWaits   = 20 * time.Second
+)
+
+// kill ends the build with this id, if it is the one running here and still working: its context
+// cancelled — which kills each command's process group — and the containers it started removed by
+// their label, once at once and again after the build has ended, so one created while it was being
+// killed is not left. The build's own goroutine announces it failed, killed by hand, and settles the
+// ask so it is not redelivered; the answer says whether that happened.
+func (h *holder) kill(id string) (string, error) {
+	h.mu.Lock()
+	r := h.running
+	if r == nil || r.request.ID != id {
+		doing := "nothing"
+		if r != nil {
+			doing = r.request.ID
+		}
+		h.mu.Unlock()
+		return "", fmt.Errorf("%s is not building %s; it is building %s. `queue` says where an ask is", h.on, id, doing)
+	}
+	if r.returned {
+		h.mu.Unlock()
+		return "", fmt.Errorf("%s on %s has already ended on its own and is saying how; `builds` shows it", id, h.on)
+	}
+	r.killed = true
+	cancel, done := r.cancel, r.done
+	h.mu.Unlock()
+
+	cancel()
+	removed, removeErr := h.removeContainers(id)
+	ended := false
+	select {
+	case <-done:
+		ended = true
+	case <-time.After(killWaits):
+	}
+	again, againErr := h.removeContainers(id)
+	removed += again
+	if removeErr == nil {
+		removeErr = againErr
+	}
+	containers := fmt.Sprintf("%d container(s) it started removed", removed)
+	if removeErr != nil {
+		containers = "its containers could not all be listed or removed: " + removeErr.Error()
+	}
+	if !ended {
+		return fmt.Sprintf("killed %s on %s: %s; the build has not finished ending yet — `builds` says when "+
+			"its outcome is in", id, h.on, containers), nil
+	}
+	h.mu.Lock()
+	outcome, announced := r.outcome, r.announced
+	h.mu.Unlock()
+	if !announced {
+		return fmt.Sprintf("killed %s (%s) on %s: its commands ended and %s, and its outcome could not be "+
+			"announced — the ask is not settled and will be handed out again", id, r.request.Repository, h.on,
+			containers), nil
+	}
+	return fmt.Sprintf("killed %s (%s) on %s: its commands ended, %s, and its outcome announced as failed, %s — "+
+		"settled, so it is not handed to another machine", id, r.request.Repository, h.on, containers, outcome), nil
+}
+
+// removeContainers is one pass of removing what the build left, bounded.
+func (h *holder) removeContainers(id string) (int, error) {
+	cleanup, stop := context.WithTimeout(context.Background(), killRemoves)
+	defer stop()
+	return builder.RemoveContainersOf(cleanup, h.remove, id)
+}
+
+// returned records that the build's work ended, and says whether a kill came first — only then is
+// the build killed; an error it ended with on its own is its own outcome.
+func (h *holder) returned(r *running) bool {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	r.returned = true
+	return r.killed
+}
+
+// said records the outcome announced, and whether it went out, for a kill to answer with.
+func (h *holder) said(r *running, outcome string, announced bool) {
+	h.mu.Lock()
+	r.outcome, r.announced = outcome, announced
+	h.mu.Unlock()
+}
+
+// handlers are the seat's verbs, as this machine answers them.
+func (h *holder) handlers() map[string]link.ToolHandler {
+	return map[string]link.ToolHandler{
+		"current": func(context.Context, json.RawMessage) (any, error) { return h.current(), nil },
+		"kill": func(_ context.Context, raw json.RawMessage) (any, error) {
+			var args struct {
+				ID string `json:"id"`
+			}
+			if err := json.Unmarshal(raw, &args); err != nil || strings.TrimSpace(args.ID) == "" {
+				return nil, errors.New("kill needs the build's id")
+			}
+			said, err := h.kill(strings.TrimSpace(args.ID))
+			if err != nil {
+				return nil, err
+			}
+			return map[string]any{"said": said}, nil
+		},
+		"pause": func(context.Context, json.RawMessage) (any, error) {
+			if err := h.setPaused(true); err != nil {
+				return nil, err
+			}
+			said := h.on + " is paused: it takes no new build until resumed"
+			if c := h.current(); c.Running != nil {
+				said += "; " + c.Running.ID + " runs on and finishes"
+			}
+			return map[string]any{"said": said, "paused": true}, nil
+		},
+		"resume": func(context.Context, json.RawMessage) (any, error) {
+			if err := h.setPaused(false); err != nil {
+				return nil, err
+			}
+			return map[string]any{"said": h.on + " takes builds again", "paused": false}, nil
+		},
+	}
+}
+
+func orNothing(s string) string {
+	if s == "" {
+		return "its start"
+	}
+	return s
+}
+
+// plainRun runs a command outside any build: no line reaches a build's log, because the build whose
+// log it would be is the one being ended.
+func plainRun(ctx context.Context, dir, name string, args ...string) (string, error) {
+	cmd := exec.CommandContext(ctx, name, args...)
+	cmd.Dir = dir
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		return string(out), fmt.Errorf("%s %s: %w: %s", name, strings.Join(args, " "), err, strings.TrimSpace(string(out)))
+	}
+	return string(out), nil
+}
+
+// announcement is what this holder answers discovery with (novox/hq ADR 0195, ADR 0197): this
+// machine's verbs of the seat, in the shape every tool runtime announces a seat's verb — kind seat,
+// the module answering, the seat, scope node, the machine, its description and argument schema — so
+// the console finds `<node>/node-build-agent.kill` by searching, as it finds any seat's verb.
+//
+// One service per machine, named for the seat and identified by the machine, so the answer is this
+// machine's four verbs and nothing more. The verbs are the compiled seat row's: a build machine has no
+// store, and what it serves is what this binary was built to serve.
+func announcement(seat, module, node string) micro.Info {
+	s, _ := catalogue.SeatNamed(seat)
+	var endpoints []micro.EndpointInfo
+	for _, v := range s.Serves {
+		schema, _ := json.Marshal(v.Input)
+		endpoints = append(endpoints, micro.EndpointInfo{
+			Name:    seat + "__" + v.Name,
+			Subject: link.NodeSeatToolSubject(seat, v.Name, node),
+			// The queue group the verbs are served in, as every runtime announces its own.
+			QueueGroup: "seat." + seat,
+			Metadata: map[string]string{
+				"kind": "seat", "module": module, "tool": v.Name, "seat": seat, "scope": "node",
+				"node": node, "interchangeable": "false", "description": v.Description, "schema": string(schema),
+			},
+		})
+	}
+	return micro.Info{
+		ServiceIdentity: micro.ServiceIdentity{Name: seat, ID: node, Version: "0.1.0",
+			Metadata: map[string]string{"seat": seat, "scope": "node", "node": node, "module": module}},
+		Description: "what the build running on " + node + " is, and ending, pausing and resuming it (novox/hq ADR 0219)",
+		Endpoints:   endpoints,
+	}
+}
+
+// moduleOf is the module a credential was issued for: its user is `<node>.<module>`.
+func moduleOf(user string) string {
+	if _, module, ok := strings.Cut(user, "."); ok && module != "" {
+		return module
+	}
+	return "build-agent"
+}

cmd/mesh-builder/holder_test.go

@@ -0,0 +1,148 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A holder paused stays paused across its restart: the flag is kept in its workspace (novox/hq ADR
+// 0219), and what it says about itself follows.
+func TestAPausedHolderStaysPausedAcrossARestart(t *testing.T) {
+	workspace := t.TempDir()
+	var said []link.HolderState
+	h := newHolder("ace", link.TheBuildMachine, workspace, func(s link.HolderState) error {
+		said = append(said, s)
+		return nil
+	})
+	if h.Paused() {
+		t.Fatal("a new holder starts paused")
+	}
+	if _, err := h.handlers()["pause"](context.Background(), json.RawMessage(`{}`)); err != nil {
+		t.Fatal(err)
+	}
+	if !h.Paused() || len(said) != 1 || !said[0].Paused || said[0].On != "ace" {
+		t.Fatalf("paused: %v, said %+v", h.Paused(), said)
+	}
+	again := newHolder("ace", link.TheBuildMachine, workspace, nil)
+	if !again.Paused() {
+		t.Fatal("restarted, the holder forgot it was paused")
+	}
+	if _, err := again.handlers()["resume"](context.Background(), json.RawMessage(`{}`)); err != nil {
+		t.Fatal(err)
+	}
+	if newHolder("ace", link.TheBuildMachine, workspace, nil).Paused() {
+		t.Fatal("resumed, the holder came back paused")
+	}
+}
+
+// kill ends the build with that id — its context, which ends its commands — removes what it left by
+// label, and refuses an id it is not building.
+func TestKillEndsTheBuildRunningHereAndNoOther(t *testing.T) {
+	h := newHolder("ace", link.TheBuildMachine, t.TempDir(), nil)
+	var removed []string
+	h.remove = func(_ context.Context, _ string, name string, args ...string) (string, error) {
+		removed = append(removed, name+" "+strings.Join(args, " "))
+		if args[0] == "ps" {
+			return "c1\n", nil
+		}
+		return "", nil
+	}
+	kill := h.handlers()["kill"]
+	if _, err := kill(context.Background(), json.RawMessage(`{"id":"build-1"}`)); err == nil {
+		t.Fatal("killed a build while none ran")
+	}
+
+	building, cancel := context.WithCancel(context.Background())
+	r := h.begin(link.BuildRequest{ID: "build-1", Repository: "novox/a"}, cancel)
+	h.stepped(r, "image")
+	if c := h.current(); c.Running == nil || c.Running.ID != "build-1" || c.Running.Step != "image" {
+		t.Fatalf("current says %+v", c)
+	}
+	if _, err := kill(context.Background(), json.RawMessage(`{"id":"build-2"}`)); err == nil ||
+		!strings.Contains(err.Error(), "build-1") {
+		t.Fatalf("killed another id: %v", err)
+	}
+	// The build's own goroutine: it ends when its context does, as a build's commands do, and
+	// announces what came of it.
+	go func() {
+		<-building.Done()
+		if h.returned(r) {
+			h.said(r, link.KilledByHand, true)
+		}
+		h.end(r)
+	}()
+	answer, err := kill(context.Background(), json.RawMessage(`{"id":"build-1"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if building.Err() == nil {
+		t.Fatal("the build's context was not cancelled")
+	}
+	if !r.killed {
+		t.Error("the build is not marked killed, so it would be announced as an ordinary failure")
+	}
+	said := answer.(map[string]any)["said"].(string)
+	if !strings.Contains(said, "2 container(s)") || !strings.Contains(said, "announced as failed") || !strings.Contains(said, link.KilledByHand) {
+		t.Errorf("kill said %q", said)
+	}
+	// Removed at the kill and again once the build had ended: a container made in between is caught.
+	if len(removed) != 4 || !strings.Contains(removed[0], "label=mesh.build=build-1") || !strings.Contains(removed[2], "label=mesh.build=build-1") {
+		t.Errorf("removed %v", removed)
+	}
+	if c := h.current(); c.Running != nil {
+		t.Errorf("after the kill current says %+v", c)
+	}
+}
+
+// A holder announces this machine's verbs of the seat as the console reads a seat's verb, and no more.
+func TestAHolderAnnouncesItsMachinesVerbsForTheConsole(t *testing.T) {
+	info := announcement(link.TheBuildMachine, moduleOf("ace.build-agent"), "ace")
+	if info.Name != "node-build-agent" || info.ID != "ace" || len(info.Endpoints) != 4 {
+		t.Fatalf("announced %s/%s with %d endpoints", info.Name, info.ID, len(info.Endpoints))
+	}
+	kill := info.Endpoints[1]
+	md := kill.Metadata
+	if kill.Subject != "mesh.seat.node-build-agent.tool.kill.ace" || kill.QueueGroup != "seat.node-build-agent" || md["kind"] != "seat" || md["seat"] != "node-build-agent" ||
+		md["scope"] != "node" || md["node"] != "ace" || md["tool"] != "kill" || md["module"] != "build-agent" ||
+		!strings.Contains(md["description"], "killed by hand") || !strings.Contains(md["schema"], `"id"`) {
+		t.Fatalf("kill is announced as %+v", kill)
+	}
+	body, err := json.Marshal(info)
+	if err != nil || len(body) > 8*1024 {
+		t.Fatalf("the answer is %d bytes (%v)", len(body), err)
+	}
+}
+
+// A build that ended on its own as the kill arrived says what it did: the kill is refused, and its
+// own error is its outcome. One whose outcome could not be announced is not said to be settled.
+func TestAKillArrivingAfterTheBuildEndedIsRefusedAndAnUnannouncedKillSaysSo(t *testing.T) {
+	h := newHolder("ace", link.TheBuildMachine, t.TempDir(), nil)
+	h.remove = func(context.Context, string, string, ...string) (string, error) { return "", nil }
+	_, cancel := context.WithCancel(context.Background())
+	r := h.begin(link.BuildRequest{ID: "build-1"}, cancel)
+	if killed := h.returned(r); killed {
+		t.Fatal("a build nobody killed reads as killed")
+	}
+	if _, err := h.kill("build-1"); err == nil || !strings.Contains(err.Error(), "ended on its own") {
+		t.Fatalf("killed a build that had ended: %v", err)
+	}
+	h.end(r)
+
+	building, cancel := context.WithCancel(context.Background())
+	r = h.begin(link.BuildRequest{ID: "build-2"}, cancel)
+	go func() {
+		<-building.Done()
+		if h.returned(r) {
+			h.said(r, link.KilledByHand, false)
+		}
+		h.end(r)
+	}()
+	said, err := h.kill("build-2")
+	if err != nil || strings.Contains(said, "announced as failed") || !strings.Contains(said, "could not be announced") {
+		t.Fatalf("kill said %q (%v)", said, err)
+	}
+}

cmd/mesh-builder/main.go

@@ -0,0 +1,569 @@
+// mesh-builder — the thing a build machine runs.
+//
+// It takes work from the mesh, turns a repository into artifacts, publishes them, and says what
+// came out. It is **not** the control plane and it is **not** the host:
+//
+//   - the control plane decides and never touches a machine. Building runs commands on one, and
+//     what the control plane may send a machine is bounded by the declaration language
+//     (novox/hq ADR 0005). "Run this build" is not in it, and widening the language so it could
+//     be would make the control plane able to run anything anywhere.
+//   - the host applies declarations and holds no opinion about what they contain. A host that
+//     also built things would need a container runtime and git, on every machine, to do something
+//     almost none of them will ever do.
+//
+// So it is a module: a program a machine runs because the mesh told it to, holding its own broker
+// credential and nothing else. Compromise of a build machine is compromise of a build machine.
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"log"
+	"net/url"
+	"os"
+	"os/signal"
+	"strings"
+	"syscall"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/builder"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// version is set at build time.
+var version = "development"
+
+func main() {
+	if err := run(); err != nil {
+		fmt.Fprintf(os.Stderr, "mesh-builder: %v\n", err)
+		os.Exit(1)
+	}
+}
+
+const usage = `mesh-builder — builds modules for the mesh
+
+It consumes build requests and answers with what it made. Nothing is listened on and nothing
+is dialled except the broker.
+
+  MESH_BROKER_FILE   a file the mesh sealed to this machine holding the same
+  MESH_REGISTRY      host:port to publish artifacts to, when the mesh has not said
+  MESH_BINDING       a file the mesh wrote saying where the artifact store is
+  MESH_PACKAGE_BINDING  a file the mesh wrote saying where the package registry is
+  MESH_NPM_TOKEN_FILE   a file the mesh sealed holding the token for it
+  MESH_NPM_REGISTRY  a package registry URL, when the mesh has not said (a person, the bootstrap)
+  MESH_NPM_TOKEN     the token for it, likewise
+  MESH_NPM_SCOPE     the scope it answers for (default: @novox)
+  MESH_WORKSPACE     where to clone and build (default: a temporary directory)
+
+It also builds one module and stops, which is how a mesh is raised — before there is a
+broker to take work from or a registry to publish into:
+
+  mesh-builder build <repository> [--path P] [--ref COMMIT] [--registry HOST:PORT]
+
+Without --registry the artifacts stay in this machine's container runtime, named by the
+digest of their own configuration. The result is printed as JSON.
+`
+
+func run() error {
+	if len(os.Args) > 1 {
+		switch os.Args[1] {
+		case "version":
+			fmt.Println(version)
+			return nil
+		case "build":
+			return buildOnce(context.Background(), os.Args[2:])
+		default:
+			fmt.Print(usage)
+			return nil
+		}
+	}
+
+	credential, err := brokerFrom()
+	if err != nil {
+		return err
+	}
+	registry, err := whereToPublish()
+	if err != nil {
+		return err
+	}
+	workspace := os.Getenv("MESH_WORKSPACE")
+	if workspace == "" {
+		workspace = os.TempDir() + "/mesh-builder"
+	}
+	// **The mesh's name for this machine, not the container's.** A build is reported to the rest
+	// of the mesh, and a report whose origin reads `104cb10e105b` names something no other module
+	// can look up. The mesh already knows the answer and has a way to say it — `${machine:name}`
+	// in the environment file this module is handed — so the hostname is only what is left when
+	// nobody said.
+	on := os.Getenv("MESH_NODE")
+	if on == "" {
+		hostname, err := os.Hostname()
+		if err != nil {
+			return fmt.Errorf("this build machine has no name: nothing said MESH_NODE and the host would not say either: %w", err)
+		}
+		on = hostname
+	}
+
+	ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM)
+	defer stop()
+
+	js, seat, err := dialFor(credential)
+	if err != nil {
+		return err
+	}
+	defer js.Close()
+
+	// **This machine's holder: paused or not, and the build it is running** (novox/hq ADR 0219). The
+	// paused flag is read from the workspace before anything is taken, so a holder restarted while
+	// paused takes nothing.
+	h := newHolder(on, seat, workspace, func(state link.HolderState) error {
+		body, err := json.Marshal(state)
+		if err != nil {
+			return err
+		}
+		_, err = js.Context().Publish(link.BuildPausedOf(seat, on), body)
+		return err
+	})
+	if h.Paused() {
+		fmt.Fprintf(os.Stderr, "paused (kept in %s): taking no build until resumed\n", pausedFile(workspace))
+	}
+	machine := link.MachineOverNATSWith(js, on, seat, link.MachineOptions{Paused: h.Paused})
+	defer machine.Close()
+
+	// The seat's verbs, on this machine's subjects. Only the seat that declares them: the retired
+	// one serves none, and a subscription its holder has no grant for would be refused for ever.
+	if seat == link.TheBuildMachine {
+		stopServing, err := link.OverNATS{Conn: js.Conn()}.ServeNodeSeatTools(seat, on, h.handlers(),
+			log.New(os.Stderr, "", 0))
+		if err != nil {
+			return err
+		}
+		defer stopServing()
+		// And says so, for the console to find (novox/hq ADR 0197).
+		stopAnnouncing, err := link.OverNATS{Conn: js.Conn()}.Announce(
+			announcement(seat, moduleOf(credential.User), on), log.New(os.Stderr, "", 0))
+		if err != nil {
+			return err
+		}
+		defer stopAnnouncing()
+		go h.stateWhilePaused(ctx, time.Hour)
+	}
+
+	fmt.Fprintf(os.Stderr, "building for the mesh, publishing to %s\n", registry)
+	publisher := builder.Registry{Address: registry, Run: builder.Command}
+
+	return machine.Take(ctx, func(ctx context.Context, work link.Build) {
+		answer(ctx, publisher, on, workspace, work, h)
+	})
+}
+
+// dialFor opens this machine's link to whichever bus the mesh is on, and says which build seat it
+// holds.
+//
+// **One place chooses**, as everywhere else the bus change went (novox/hq ADR 0116 step 5): a build
+// machine told about both would take work from one and answer on the other, and every log line would
+// say it was fine.
+func dialFor(credential Credential) (*broker.JetStream, string, error) {
+	// **The credential names the bus, and there is one** (novox/hq ADR 0131, design 28 task 5.5).
+	// A credential for the mesh's bus carries user, password and fingerprint beside the address,
+	// and that is enough to dial it, pinned.
+	if !credential.onTheNewBus() {
+		return nil, "", fmt.Errorf("the credential at hand names %q, which is not the mesh's bus", credential.URL)
+	}
+	js, err := broker.DialPinned(credential.natsURL(), credential.Fingerprint)
+	if err != nil {
+		return nil, "", err
+	}
+	// **The seat this machine serves is the one its credential claims** (novox/hq ADR 0190, the
+	// handover): the mesh issues a build machine's credential naming the seat its module claims,
+	// and one binary serves the old role as `builder` and the new as `build-agent` from that alone.
+	seat := link.BuildSeatClaimed(credential.seatsClaimed())
+	fmt.Fprintf(os.Stderr, "taking build work as a holder of %s\n", seat)
+	return js, seat, nil
+}
+
+// answer does one build and says what happened, whichever way it went.
+func answer(ctx context.Context, publisher builder.Publisher, on, workspace string, work link.Build, h *holder) {
+	request := work.Request()
+
+	// **Its own context, so it can be killed alone** (novox/hq ADR 0219): cancelled by `kill`, it ends
+	// this build's commands and nothing else; the machine's own context ending — a SIGTERM — still
+	// reaches it through the parent, and that keeps today's meaning below.
+	building, cancel := context.WithCancel(ctx)
+	defer cancel()
+	var mine *running
+	if h != nil {
+		mine = h.begin(request, cancel)
+		defer h.end(mine)
+	}
+
+	// **First thing, and to stdout.** A build request that arrives and produces no visible line until
+	// it either finishes or fails is indistinguishable from one that never arrived — which cost a long
+	// diagnosis against a running mesh, chasing "the handler never fired" when the truth was only that
+	// the handler said nothing until the end.
+	fmt.Fprintf(os.Stderr, "a build request arrived for %s (%s)\n", request.Repository, request.ID)
+
+	// **Everything a build says goes two ways**: to stderr, as always, and onto the bus as the
+	// role's own events under the build's id (novox/hq ADR 0157) — so whoever asked, and anybody
+	// watching, reads the same lines this container's log holds, live, and after the fact from the
+	// stream. Said first, before anything runs, so a build that hangs is one that visibly started.
+	say := func(step, message string) {
+		fmt.Fprintf(os.Stderr, "  [%s] %s\n", step, message)
+		work.Say(step, message)
+		if mine != nil && step != "run" && step != "output" {
+			h.stepped(mine, step)
+		}
+	}
+	builder.Said = say
+	defer func() { builder.Said = nil }()
+	if err := work.Began(ctx); err != nil {
+		fmt.Fprintf(os.Stderr, "cannot say a build started: %v\n", err)
+	}
+
+	result := link.BuildResult{
+		ID: request.ID, Repository: request.Repository, Path: request.Path,
+		Ref: request.Ref, On: on, Source: request.Source, DryRun: request.DryRun,
+	}
+	what := "building " + request.Repository
+	if request.Path != "" {
+		what += " at " + request.Path
+	}
+	if request.Ref != "" {
+		what += " on " + request.Ref
+	}
+	say("build", what)
+
+	npmrc, err := packagesFrom()
+	var built builder.Result
+	if err == nil {
+		// The package-registry credential is a build input, so it is resolved before the clone: a
+		// build that could not have resolved its dependencies is refused in front of the reason, not
+		// after a clone that then fails at npm ci.
+		// Every container it starts is labelled with its id, so a kill finds what outlived the
+		// docker client (ADR 0219).
+		built, err = builder.Build(building, builder.Labelled(builder.Command, request.ID), publisher,
+			request.Repository, request.Path, request.Ref, workspace, request.Held, npmrc,
+			forgeFrom(), say, request.Seats)
+	}
+	// Only a build the kill ended: the kill came before its work did. One that finished — built, or
+	// failed on its own — in the moment the kill arrived says what it did, and the kill is refused.
+	killed := false
+	if mine != nil {
+		killed = h.returned(mine) && err != nil
+	}
+	if killed {
+		// **Killed by hand is the outcome, whatever the build was doing** (novox/hq ADR 0219): the
+		// error it ended with is the kill's consequence, not a fault of the source.
+		result.Failed = link.KilledByHand
+		say("failed", link.KilledByHand)
+	} else if err != nil {
+		// A failure is a result. A build that fails and says nothing is indistinguishable from a
+		// builder that is not running, and those want completely different responses.
+		result.Failed = err.Error()
+		say("failed", err.Error())
+	} else {
+		manifest, marshalErr := json.Marshal(built.Manifest)
+		if marshalErr != nil {
+			result.Failed = marshalErr.Error()
+		} else {
+			result.Commit = built.Commit
+			result.Manifest = manifest
+			for _, made := range built.Built {
+				result.Made = append(result.Made, link.MadeArtifact{
+					Name: made.Name, Kind: made.Kind, Reference: made.Reference,
+				})
+			}
+			result.Against = built.Against
+			for _, r := range built.Read {
+				result.Read = append(result.Read, link.ReadRepository{Repository: r.Repository, Ref: r.Ref})
+			}
+			say("built", built.Manifest.Module+" from "+short(built.Commit))
+		}
+	}
+
+	// A killed build is announced on a context of its own: the build's was the one cancelled, and the
+	// outcome must go out and the ask be settled — acknowledged, never redelivered to another machine
+	// to be built again. A machine being stopped is the other case and keeps its meaning: the
+	// parent's context is gone, nothing is announced or settled, and the ask is redelivered.
+	announcing := ctx
+	if killed {
+		fresh, stop := context.WithTimeout(context.Background(), 30*time.Second)
+		defer stop()
+		announcing = fresh
+	}
+	announceErr := work.Announce(announcing, result)
+	if mine != nil {
+		h.said(mine, result.Failed, announceErr == nil)
+	}
+	if err := announceErr; err != nil {
+		// Said, not fatal: the build happened. A build reported as failed because announcing it
+		// failed is a lie about work that was done — and the request stays unsettled below only if
+		// nothing was said at all, so another machine can try.
+		fmt.Fprintf(os.Stderr, "cannot say what came of a build: %v\n", err)
+		return
+	}
+
+	// Settled only once the outcome is away, so a machine that dies before answering leaves the work
+	// for another rather than losing it.
+	if err := work.Done(); err != nil {
+		fmt.Fprintf(os.Stderr, "the outcome is away and the request could not be settled: %v\n", err)
+	}
+}
+
+// packagesFrom is where a build resolves the mesh's own published packages — the SDK above all
+// (novox/hq ADR 0076, issue 053).
+//
+// Preferably from the mesh: a package-registry binding names the endpoint the way the artifact
+// store's binding does, and a sealed token file the credential the way the broker's does. The
+// environment variables remain for a builder run by a person, and for the bootstrap, where there is
+// no registry yet — there the result is disabled and a build that needs no mesh-published dependency
+// builds anyway.
+func packagesFrom() (builder.Npmrc, error) {
+	scope := strings.TrimSpace(os.Getenv("MESH_NPM_SCOPE"))
+	if scope == "" {
+		scope = "@novox"
+	}
+
+	registry := strings.TrimSpace(os.Getenv("MESH_NPM_REGISTRY"))
+	var username string
+	if path := strings.TrimSpace(os.Getenv("MESH_PACKAGE_BINDING")); path != "" {
+		raw, err := os.ReadFile(path)
+		if err != nil {
+			return builder.Npmrc{}, fmt.Errorf("cannot read what the mesh said about the package registry: %w", err)
+		}
+		var told struct {
+			From   string         `json:"from"`
+			At     string         `json:"at"`
+			As     string         `json:"as"`
+			Serves map[string]any `json:"serves"`
+		}
+		if err := json.Unmarshal(raw, &told); err != nil {
+			return builder.Npmrc{}, fmt.Errorf("%s is not a binding: %w", path, err)
+		}
+		if told.At == "" {
+			return builder.Npmrc{}, fmt.Errorf(
+				"%s says the package registry is on %q and gives no address for it", path, told.From)
+		}
+		// Composed from what the provider serves, so nothing here knows gitea's URL shape from
+		// another registry's: it states its port, the path its registry answers on, and the scheme.
+		scheme := "https"
+		if s, ok := told.Serves["scheme"]; ok {
+			scheme = fmt.Sprintf("%v", s)
+		}
+		port, ok := told.Serves["port"]
+		if !ok {
+			return builder.Npmrc{}, fmt.Errorf("%s says nothing about which port the package registry answers on", path)
+		}
+		npmPath, ok := told.Serves["npm-path"]
+		if !ok {
+			return builder.Npmrc{}, fmt.Errorf("%s says nothing about the path the package registry answers on", path)
+		}
+		registry = fmt.Sprintf("%s://%s:%v%v", scheme, told.At, port, npmPath)
+		username = told.As
+	}
+
+	// The credential the mesh sealed to this machine. The mesh authenticates the ordinary way — a
+	// generated password the provider only applies (novox/hq ADR 0048) — so with a username this is
+	// a password (basic auth); without one it is a bearer token a provider minted.
+	secret := strings.TrimSpace(os.Getenv("MESH_NPM_TOKEN"))
+	if path := strings.TrimSpace(os.Getenv("MESH_NPM_TOKEN_FILE")); path != "" {
+		raw, err := os.ReadFile(path)
+		if err != nil {
+			return builder.Npmrc{}, fmt.Errorf("cannot read this builder's package-registry credential: %w", err)
+		}
+		secret = strings.TrimSpace(string(raw))
+	}
+	if u := strings.TrimSpace(os.Getenv("MESH_NPM_USER")); u != "" {
+		username = u
+	}
+
+	if registry == "" && secret == "" {
+		return builder.Npmrc{}, nil
+	}
+	if username != "" {
+		return builder.Npmrc{Scope: scope, Registry: registry, Username: username, Password: secret}, nil
+	}
+	return builder.Npmrc{Scope: scope, Registry: registry, Token: secret}, nil
+}
+
+// forgeFrom is the git credential this builder may offer a clone, composed from the same binding
+// and sealed secret its package-registry half already reads: the forge that answers npm is the
+// forge that hosts the repositories, and its provisioner applies one password to one user for
+// both. Anything missing means no credential, and every clone stays anonymous — which is all a
+// mesh of public repositories ever needs.
+//
+// The URL names the binding's own address — the machine the mesh says the forge is on — so a
+// private repository is registered and built by that address, and a clone of anything else is
+// never shown this credential (git's credential store matches the whole origin).
+func forgeFrom() builder.GitCredential {
+	path := strings.TrimSpace(os.Getenv("MESH_PACKAGE_BINDING"))
+	if path == "" {
+		return builder.GitCredential{}
+	}
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		return builder.GitCredential{}
+	}
+	var told struct {
+		At     string         `json:"at"`
+		As     string         `json:"as"`
+		Serves map[string]any `json:"serves"`
+	}
+	if err := json.Unmarshal(raw, &told); err != nil || told.At == "" || told.As == "" {
+		return builder.GitCredential{}
+	}
+	secret := strings.TrimSpace(os.Getenv("MESH_NPM_TOKEN"))
+	if file := strings.TrimSpace(os.Getenv("MESH_NPM_TOKEN_FILE")); file != "" {
+		if raw, err := os.ReadFile(file); err == nil {
+			secret = strings.TrimSpace(string(raw))
+		}
+	}
+	if secret == "" {
+		return builder.GitCredential{}
+	}
+	scheme := "https"
+	if s, ok := told.Serves["scheme"]; ok {
+		scheme = fmt.Sprintf("%v", s)
+	}
+	host := told.At
+	if port, ok := told.Serves["port"]; ok {
+		host = fmt.Sprintf("%s:%v", told.At, port)
+	}
+	made := url.URL{Scheme: scheme, User: url.UserPassword(told.As, secret), Host: host}
+	return builder.GitCredential{URL: made.String()}
+}
+
+func short(commit string) string {
+	if len(commit) > 8 {
+		return commit[:8]
+	}
+	return commit
+}
+
+// whereToPublish is the artifact store this builder uses.
+//
+// **Preferably from the mesh.** A builder that is a module requires an artifact store, and the
+// mesh writes it a file saying which machine answers that and on what port — the same binding any
+// consumer of any provision gets. Reading it means the address is not a setting somebody keeps in
+// step by hand, and moving the store is an ordinary reassignment rather than an edit on every
+// build machine.
+//
+// The environment variable remains for a builder run by a person, which is how this started and
+// how it is still run while being developed.
+func whereToPublish() (string, error) {
+	binding := strings.TrimSpace(os.Getenv("MESH_BINDING"))
+	if binding == "" {
+		registry := strings.TrimSpace(os.Getenv("MESH_REGISTRY"))
+		if registry == "" {
+			return "", fmt.Errorf("neither MESH_BINDING nor MESH_REGISTRY: a built artifact " +
+				"nobody can fetch is not built")
+		}
+		return registry, nil
+	}
+
+	raw, err := os.ReadFile(binding)
+	if err != nil {
+		return "", fmt.Errorf("cannot read what the mesh said about the artifact store: %w", err)
+	}
+	var told struct {
+		From   string         `json:"from"`
+		At     string         `json:"at"`
+		Serves map[string]any `json:"serves"`
+	}
+	if err := json.Unmarshal(raw, &told); err != nil {
+		return "", fmt.Errorf("%s is not a binding: %w", binding, err)
+	}
+	if told.At == "" {
+		// The provider is not on the private network, so there is no name to reach it by. Said
+		// rather than falling back to the machine's own name, which would publish to a store on
+		// the wrong machine and be found out much later.
+		return "", fmt.Errorf(
+			"%s says the artifact store is on %q and gives no address for it", binding, told.From)
+	}
+	port, ok := told.Serves["port"]
+	if !ok {
+		return "", fmt.Errorf("%s says nothing about which port the artifact store answers on",
+			binding)
+	}
+	return fmt.Sprintf("%s:%v", told.At, port), nil
+}
+
+// brokerFrom is where this builder connects, and with what.
+//
+// **Preferably from a file the mesh sealed to this machine.** A builder that is a module is given
+// its credential the way every other module is given one: generated or accepted centrally, sealed
+// to the machine, written by the host. Putting it in an environment variable instead would mean
+// the one copy that matters passing through a terminal and a process listing.
+//
+// The variable remains for a builder run by a person.
+func brokerFrom() (Credential, error) {
+	if path := strings.TrimSpace(os.Getenv("MESH_BROKER_FILE")); path != "" {
+		raw, err := os.ReadFile(path)
+		if err != nil {
+			return Credential{}, fmt.Errorf("cannot read this builder's credential: %w", err)
+		}
+		said := strings.TrimSpace(string(raw))
+		if said == "" {
+			// An empty credential file is a machine that will connect as nobody and be refused,
+			// with the reason three layers away.
+			return Credential{}, fmt.Errorf("%s is empty, so this builder has no credential", path)
+		}
+		var held Credential
+		if err := json.Unmarshal([]byte(said), &held); err == nil && held.URL != "" {
+			return held, nil
+		}
+		// A file holding only a URL, which is what a person writing one by hand produces. The
+		// broker is then verified against whatever this machine already trusts.
+		return Credential{URL: said}, nil
+	}
+	return Credential{}, fmt.Errorf(
+		"no MESH_BROKER_FILE: a build machine with no credential for the bus has nothing to build")
+}
+
+// Credential is what a build machine is given so it can reach the broker.
+//
+// Two things, because reaching a broker over TLS needs both: who to connect as, and what to check
+// the certificate against. A mesh's broker presents a certificate of the mesh's own, which is in
+// no public trust store, so a URL alone can only connect to a broker somebody else vouches for.
+//
+// **The same shape a node gets, for the same reason** (novox/hq ADR 0004): the fingerprint travels
+// out of band — here, sealed with the credential — and the endpoint is verified once at connect.
+type Credential struct {
+	URL         string `json:"url"`
+	Fingerprint string `json:"fingerprint,omitempty"`
+	// User and Password ride beside the address on the bus being built (design 25): a credential
+	// embedded in a URL leaks into every log line that prints a connection, so the mesh seals them
+	// as two fields and this machine joins them once, here, to dial.
+	User     string `json:"user,omitempty"`
+	Password string `json:"password,omitempty"`
+	// Claims are the seats the module this credential was issued for claims, as the mesh writes
+	// them beside the credential (novox/hq ADR 0159). The first is the build role this machine
+	// serves; a credential naming none is from before claims travelled in it.
+	Claims []struct {
+		Seat string `json:"seat"`
+	} `json:"claims,omitempty"`
+}
+
+// seatsClaimed is the seats the credential names, in order.
+func (c Credential) seatsClaimed() []string {
+	out := make([]string, 0, len(c.Claims))
+	for _, claim := range c.Claims {
+		out = append(out, claim.Seat)
+	}
+	return out
+}
+
+// onTheNewBus is whether a credential is for the bus being built: its address says so, and the
+// mesh only ever seals such a credential with the user and password beside it.
+func (c Credential) onTheNewBus() bool { return strings.HasPrefix(strings.TrimSpace(c.URL), "nats://") }
+
+// natsURL is the address with this machine's credential in it, for the one dial that needs it.
+func (c Credential) natsURL() string {
+	rest := strings.TrimPrefix(strings.TrimSpace(c.URL), "nats://")
+	if c.User == "" {
+		return "nats://" + rest
+	}
+	return "nats://" + c.User + ":" + c.Password + "@" + rest
+}

cmd/mesh-builder/once.go

@@ -0,0 +1,181 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"os"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/builder"
+)
+
+// buildOnce is the builder doing one build and stopping, with no broker and no mesh.
+//
+// **This is how a mesh is raised** (novox/hq ADR 0073). The installer carries this program and runs
+// it once, before anything else exists, to produce the control plane from the same repository and
+// path that every later rebuild of it will use. What raises the mesh is therefore the same thing
+// that maintains it — not a second mechanism that has to be kept in step with the first and is
+// exercised once per new mesh, which is how often enough to rot.
+//
+// It takes no work from a queue and answers nobody: there is no broker yet, and the only thing
+// waiting for the answer is the installer that started it. So the result goes to standard output as
+// JSON, which is what the installer reads.
+//
+// With no --registry it publishes nowhere and the image stays in this machine's container runtime,
+// named by the digest of its own configuration (see builder.Local). That is the genesis case. With
+// one, it publishes as it always does — the same command is how an operator builds a module by hand
+// on a mesh that already exists.
+func buildOnce(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("build", flag.ContinueOnError)
+	path := set.String("path", "", "the module's directory inside the repository (default: its root)")
+	ref := set.String("ref", "", "the commit to build; a branch is a moving target somebody else controls")
+	registry := set.String("registry", "",
+		"host:port to publish to. Without it the artifacts stay in this machine's container runtime, which is the genesis case")
+	workspace := set.String("workspace", "", "where to clone and build (default: a temporary directory)")
+	on := multiple(set, "on",
+		"a base this build stands on, as <module>/<artifact>=<reference>. Repeatable")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 1 {
+		return errors.New("mesh-builder build <repository> [--path P] [--ref COMMIT] [--registry HOST:PORT]")
+	}
+	repository := positionals[0]
+
+	where := *workspace
+	if where == "" {
+		where = os.TempDir() + "/mesh-builder-once"
+	}
+
+	// What this build may stand on, said on the command line because there is no mesh to ask.
+	//
+	// **This is the genesis case and it is meant to be awkward.** On a running mesh the control
+	// plane answers this, because only it knows what this mesh holds. Here nothing has been built
+	// yet, so whoever runs this says what to use — and for the control plane, which is the one
+	// module raised before anything else exists, the answer is ordinarily nothing at all.
+	bases := map[string]string{}
+	for _, pair := range *on {
+		key, reference, found := strings.Cut(pair, "=")
+		if !found || key == "" || reference == "" {
+			return fmt.Errorf(
+				"--on takes <module>/<artifact>=<reference>, and %q is not that", pair)
+		}
+		bases[key] = reference
+	}
+
+	// Local unless told otherwise, because the moment this exists for has nowhere to publish. A
+	// default pointing at a registry would mean genesis failing at a push to something that is not
+	// there yet, one step away from the thing that could explain it.
+	var publisher builder.Publisher = builder.Local{Run: builder.Command}
+	if *registry != "" {
+		publisher = builder.Registry{Address: *registry, Run: builder.Command}
+	}
+
+	fmt.Fprintf(os.Stderr, "building %s", repository)
+	if *path != "" {
+		fmt.Fprintf(os.Stderr, " at %s", *path)
+	}
+	if *ref != "" {
+		fmt.Fprintf(os.Stderr, " at %s", *ref)
+	}
+	fmt.Fprintln(os.Stderr)
+
+	npmrc, err := packagesFrom()
+	if err != nil {
+		return err
+	}
+	built, buildErr := builder.Build(ctx, builder.Command, publisher, repository, *path, *ref, where, bases, npmrc,
+		forgeFrom(),
+		func(step, message string) { fmt.Fprintf(os.Stderr, "  [%s] %s\n", step, message) })
+	if buildErr != nil {
+		return buildErr
+	}
+
+	// To standard output, and everything else to standard error, so the caller can read this
+	// without having to separate it from progress.
+	out := onceResult{
+		Module:     built.Manifest.Module,
+		Commit:     built.Commit,
+		Repository: repository,
+		Path:       *path,
+		Ref:        *ref,
+		Manifest:   built.Manifest,
+		Against:    built.Against,
+	}
+	for _, r := range built.Read {
+		out.Read = append(out.Read, readRepository{Repository: r.Repository, Ref: r.Ref})
+	}
+	for _, made := range built.Built {
+		out.Made = append(out.Made, madeArtifact{Name: made.Name, Kind: made.Kind, Reference: made.Reference})
+	}
+	body, err := json.MarshalIndent(out, "", "  ")
+	if err != nil {
+		return err
+	}
+	fmt.Println(string(body))
+	fmt.Fprintf(os.Stderr, "  built %s from %s\n", built.Manifest.Module, short(built.Commit))
+	return nil
+}
+
+// onceResult is what one build reports to whoever started it.
+//
+// The same fields the mesh records for a build, so a reader comparing a genesis build against an
+// ordinary one is comparing the same thing said the same way.
+type onceResult struct {
+	Module     string           `json:"module"`
+	Commit     string           `json:"commit"`
+	Repository string           `json:"repository"`
+	Path       string           `json:"path,omitempty"`
+	Ref        string           `json:"ref,omitempty"`
+	Manifest   any              `json:"manifest"`
+	Made       []madeArtifact   `json:"made"`
+	Against    []string         `json:"against,omitempty"`
+	Read       []readRepository `json:"read,omitempty"`
+}
+
+// readRepository is a repository this build read source from besides the module's own.
+type readRepository struct {
+	Repository string `json:"repository"`
+	Ref        string `json:"ref,omitempty"`
+}
+
+type madeArtifact struct {
+	Name      string `json:"name"`
+	Kind      string `json:"kind"`
+	Reference string `json:"reference"`
+}
+
+// parseAround lets flags appear on either side of the repository, because a person writing this by
+// hand will put them wherever reads best and the standard parser stops at the first thing that is
+// not a flag. The same helper the control plane's commands use, for the same reason.
+func parseAround(set *flag.FlagSet, args []string) ([]string, error) {
+	var positionals []string
+	rest := args
+	for {
+		if err := set.Parse(rest); err != nil {
+			return nil, err
+		}
+		rest = set.Args()
+		if len(rest) == 0 {
+			return positionals, nil
+		}
+		positionals = append(positionals, rest[0])
+		rest = rest[1:]
+	}
+}
+
+// multiple is a flag that may be given more than once.
+type repeated []string
+
+func (r *repeated) String() string     { return strings.Join(*r, ", ") }
+func (r *repeated) Set(v string) error { *r = append(*r, v); return nil }
+
+func multiple(set *flag.FlagSet, name, usage string) *[]string {
+	var values repeated
+	set.Var(&values, name, usage)
+	return (*[]string)(&values)
+}

cmd/mesh-builder/seat_test.go

@@ -0,0 +1,27 @@
+package main
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// The seat a build machine serves comes from its credential (novox/hq ADR 0190 handover).
+func TestTheCredentialSaysWhichBuildRoleThisMachineServes(t *testing.T) {
+	var held Credential
+	if err := json.Unmarshal([]byte(`{"url":"nats://bus:4222","user":"anchor.builder","password":"x",
+		"claims":[{"seat":"mesh-build-machine","scope":"mesh","serves":[]}]}`), &held); err != nil {
+		t.Fatal(err)
+	}
+	if got := link.BuildSeatClaimed(held.seatsClaimed()); got != "mesh-build-machine" {
+		t.Errorf("the old builder's credential serves %q", got)
+	}
+	var bare Credential
+	if err := json.Unmarshal([]byte(`{"url":"nats://bus:4222","user":"anchor.build-agent","password":"x"}`), &bare); err != nil {
+		t.Fatal(err)
+	}
+	if got := link.BuildSeatClaimed(bare.seatsClaimed()); got != link.TheBuildMachine {
+		t.Errorf("a credential without claims serves %q, want %s", got, link.TheBuildMachine)
+	}
+}

cmd/mesh-builder/stdout_test.go

@@ -0,0 +1,44 @@
+package main
+
+import (
+	"os"
+	"strings"
+	"testing"
+)
+
+// **The genesis path writes its result to stdout and nothing else there.** The installer captures
+// stdout alone and parses it as JSON; one stray log line makes that fail with "invalid character"
+// — exactly what a builder fmt.Printf caused, breaking a build that had worked. This keeps
+// diagnostics on stderr so a future print cannot repeat it silently.
+func TestBuilderDiagnosticsStayOffStdout(t *testing.T) {
+	allowed := map[string]bool{
+		"string(body)": true, // once.go: the result JSON, which IS stdout
+		"version)":     true, // --version
+		`"stopping")`:  true, // the loop.s shutdown line
+		"usage)":       true, // --help text, for a human
+	}
+	for _, file := range []string{"once.go", "main.go"} {
+		src, err := os.ReadFile(file)
+		if err != nil {
+			t.Fatal(err)
+		}
+		for i, raw := range strings.Split(string(src), "\n") {
+			line := strings.TrimSpace(raw)
+			if !strings.HasPrefix(line, "fmt.Printf(") &&
+				!strings.HasPrefix(line, "fmt.Println(") &&
+				!strings.HasPrefix(line, "fmt.Print(") {
+				continue
+			}
+			ok := false
+			for token := range allowed {
+				if strings.Contains(line, token) {
+					ok = true
+				}
+			}
+			if !ok {
+				t.Errorf("%s:%d writes a diagnostic to stdout, which the installer parses as JSON:\n  %s",
+					file, i+1, line)
+			}
+		}
+	}
+}

cmd/mesh-builder/where_test.go

@@ -0,0 +1,225 @@
+package main
+
+import (
+	"crypto/ed25519"
+	"crypto/rand"
+	"crypto/sha256"
+	"crypto/tls"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/hex"
+	"math/big"
+	"net"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// Where a builder publishes.
+//
+// Preferably from the mesh: a builder that is a module requires an artifact store, and the mesh
+// writes it a binding saying which machine answers and on what port. Reading it means the address
+// is not a setting somebody keeps in step by hand.
+
+func binding(t *testing.T, body string) string {
+	t.Helper()
+	path := filepath.Join(t.TempDir(), "artifact-store.json")
+	if err := os.WriteFile(path, []byte(body), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	return path
+}
+
+func TestTheMeshSaysWhereToPublish(t *testing.T) {
+	t.Setenv("MESH_BINDING", binding(t, `{"binding":1,"provision":"artifact-store",
+		"from":"anchor","at":"anchor.internal","serves":{"port":5000,"scheme":"http"}}`))
+	where, err := whereToPublish()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if where != "anchor.internal:5000" {
+		t.Fatalf("got %q", where)
+	}
+}
+
+func TestABindingWithNoAddressIsRefused(t *testing.T) {
+	// The provider is not on the private network, so there is no name to reach it by. Falling
+	// back to anything would publish to a store on the wrong machine and be found out much later.
+	t.Setenv("MESH_BINDING", binding(t, `{"binding":1,"from":"anchor","serves":{"port":5000}}`))
+	_, err := whereToPublish()
+	if err == nil {
+		t.Fatal("a binding with nowhere to reach was accepted")
+	}
+	if !strings.Contains(err.Error(), "anchor") {
+		t.Fatalf("the failure does not name the machine: %v", err)
+	}
+}
+
+func TestABindingWithNoPortIsRefused(t *testing.T) {
+	t.Setenv("MESH_BINDING", binding(t, `{"binding":1,"from":"a","at":"a.internal","serves":{}}`))
+	if _, err := whereToPublish(); err == nil {
+		t.Fatal("a binding saying nothing about a port was accepted")
+	}
+}
+
+func TestTheVariableStillWorksForABuilderRunByAPerson(t *testing.T) {
+	// Which is how this started and how it is still run while being developed.
+	t.Setenv("MESH_BINDING", "")
+	t.Setenv("MESH_REGISTRY", "127.0.0.1:5000")
+	where, err := whereToPublish()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if where != "127.0.0.1:5000" {
+		t.Fatalf("got %q", where)
+	}
+}
+
+func TestNeitherIsRefusedRatherThanGuessed(t *testing.T) {
+	t.Setenv("MESH_BINDING", "")
+	t.Setenv("MESH_REGISTRY", "")
+	if _, err := whereToPublish(); err == nil {
+		t.Fatal("a builder with nowhere to publish reported somewhere")
+	}
+}
+
+func TestTheCredentialComesFromAFileTheMeshSealed(t *testing.T) {
+	// A builder that is a module is given its credential the way every module is: sealed to the
+	// machine and written by the host. An environment variable instead would put the one copy
+	// that matters through a terminal and a process listing.
+	path := filepath.Join(t.TempDir(), "broker")
+	if err := os.WriteFile(path, []byte("amqps://a-builder:secret@broker.internal:5671/\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("MESH_BROKER_FILE", path)
+	t.Setenv("MESH_BROKER_AMQP", "amqp://should-not-be-used@nowhere/")
+
+	got, err := brokerFrom()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got.URL != "amqps://a-builder:secret@broker.internal:5671/" {
+		t.Fatalf("got %q", got.URL)
+	}
+}
+
+// The credential the mesh seals carries what to check the broker's certificate against, because a
+// mesh's broker presents a certificate of the mesh's own and no public trust store has it. A URL
+// alone can only reach a broker somebody else vouches for.
+func TestTheSealedCredentialCarriesWhatVerifiesTheBroker(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "broker")
+	if err := os.WriteFile(path, []byte(
+		`{"url":"amqps://a-builder:secret@broker.internal:5671/","fingerprint":"abc123"}`),
+		0o600); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("MESH_BROKER_FILE", path)
+	t.Setenv("MESH_BROKER_AMQP", "")
+
+	got, err := brokerFrom()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got.URL != "amqps://a-builder:secret@broker.internal:5671/" {
+		t.Fatalf("the url was lost: %q", got.URL)
+	}
+	if got.Fingerprint != "abc123" {
+		t.Fatal("the builder was given nothing to check the broker against, so it can only " +
+			"connect to a broker some public authority vouches for")
+	}
+}
+
+func TestAnEmptyCredentialFileIsRefused(t *testing.T) {
+	// Otherwise the builder connects as nobody and is refused, with the reason three layers away.
+	path := filepath.Join(t.TempDir(), "broker")
+	if err := os.WriteFile(path, []byte("\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("MESH_BROKER_FILE", path)
+	t.Setenv("MESH_BROKER_AMQP", "")
+	if _, err := brokerFrom(); err == nil {
+		t.Fatal("an empty credential was accepted")
+	}
+}
+
+func TestABuilderWithNoCredentialAtAllSaysSo(t *testing.T) {
+	t.Setenv("MESH_BROKER_FILE", "")
+	t.Setenv("MESH_BROKER_AMQP", "")
+	if _, err := brokerFrom(); err == nil {
+		t.Fatal("a builder with no broker reported one")
+	}
+}
+
+// The pin is compared in the spelling the mesh writes it.
+//
+// A bare digest against a written fingerprint never matches, and the failure is indistinguishable
+// from being pointed at the wrong broker — which is the one thing this check exists to report
+// truthfully. It cost a lab run.
+func TestThePinIsComparedInTheSpellingTheMeshWritesIt(t *testing.T) {
+	certificate, key := aServerCertificate(t)
+	der := certificate.Certificate[0]
+	sum := sha256.Sum256(der)
+	written := "sha256:" + hex.EncodeToString(sum[:])
+
+	listener, err := tls.Listen("tcp", "127.0.0.1:0", &tls.Config{
+		Certificates: []tls.Certificate{certificate}, MinVersion: tls.VersionTLS12,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer listener.Close()
+	go func() {
+		for {
+			conn, err := listener.Accept()
+			if err != nil {
+				return
+			}
+			_ = conn.(*tls.Conn).Handshake()
+			conn.Close()
+		}
+	}()
+	_ = key
+
+	// The pin the mesh wrote must be accepted.
+	if err := handshakeWith(listener.Addr().String(), written); err != nil {
+		t.Fatalf("the broker this builder was told about was refused: %v", err)
+	}
+	// And a different one refused, or the check reports nothing.
+	other := "sha256:" + strings.Repeat("ab", 32)
+	if err := handshakeWith(listener.Addr().String(), other); err == nil {
+		t.Fatal("a broker this builder was not told about was accepted")
+	}
+}
+
+// handshakeWith runs the pin check the builder dials with against an address.
+func handshakeWith(address, pin string) error {
+	conn, err := tls.Dial("tcp", address, broker.PinnedToFingerprint(pin))
+	if err != nil {
+		return err
+	}
+	return conn.Close()
+}
+
+func aServerCertificate(t *testing.T) (tls.Certificate, ed25519.PrivateKey) {
+	t.Helper()
+	public, private, err := ed25519.GenerateKey(rand.Reader)
+	if err != nil {
+		t.Fatal(err)
+	}
+	template := &x509.Certificate{
+		SerialNumber: big.NewInt(1),
+		Subject:      pkix.Name{CommonName: "a broker"},
+		NotBefore:    time.Now().Add(-time.Hour),
+		NotAfter:     time.Now().Add(time.Hour),
+		IPAddresses:  []net.IP{net.ParseIP("127.0.0.1")},
+	}
+	der, err := x509.CreateCertificate(rand.Reader, template, template, public, private)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return tls.Certificate{Certificate: [][]byte{der}, PrivateKey: private}, private
+}

cmd/mesh-controller/acts.go

@@ -0,0 +1,317 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"github.com/novox/mesh-controller/internal/broker"
+	"slices"
+	"sort"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// The things the mesh can be asked to do, separated from how it was asked.
+//
+// **A surface is an adapter with no decisions in it** (novox/hq ADR 0035). The command line and
+// the command API call the same functions here, so an assignment refused at one is refused at the
+// other for the same reason and in the same words. The moment a surface can accept something
+// another would reject, the mesh has two answers to one question and people learn which to trust.
+//
+// Each returns what happened as text a person can read and a caller can pass on. Neither surface
+// composes its own explanation, because two explanations of one refusal drift.
+
+// assign puts a module on a node, and says at once what in the mesh no longer works out.
+//
+// The assignment is kept even when the node does not resolve: it is what a person meant, and
+// assignment is not an ordering. A consumer assigned before its provider does not resolve for as
+// long as it takes to assign the provider, and refusing the first half of a pair would make the
+// order somebody types two commands in part of the mesh's rules.
+//
+// **What is checked is the mesh, not the one machine** — because that is what the assignment
+// changes. novox/hq 04-ISSUES/017 names the shape: an action can succeed into a state its own
+// verify rejects, and it happens when the action's test is not the test the verify uses. Here the
+// action tested one node and the verify is resolution over all of them, so an assignment could be
+// reported as fine while it took a provision away from every other machine — a module offering a
+// mesh-scoped provision stops offering it the moment its own node stops resolving, and every
+// consumer elsewhere is then told *nothing in this mesh provides it*, with a remedy that names a
+// module already assigned. That was found on a four-node raise and read as a version bump breaking
+// provider recognition; it was neither the version nor the provider.
+//
+// It costs a resolution per machine. Assignment is a person typing a command, and being told which
+// machines this just blocked is worth more than the milliseconds.
+func assign(ctx context.Context, open *stores, node string, modules ...string) (string, error) {
+	if len(modules) == 0 {
+		return "", fmt.Errorf("assign %s names no module", node)
+	}
+	// Held while it is recorded, so it cannot land between a converge's preview and its flip and
+	// be taken without ever having been previewed (novox/hq ADR 0100).
+	ctx, release, err := holdNodes(ctx, open, []string{node})
+	if err != nil {
+		return "", err
+	}
+	defer release()
+	// **The one assignment refused for what the node lacks** (novox/hq ADR 0207). Everything else
+	// an assignment leaves unresolved is kept, because assignment is not an ordering; a module whose
+	// resources are applied through a seat nothing on the node holds is refused, because that order
+	// — the service manager, the package manager and the runtime before anything that installs,
+	// runs or contains — is the mesh's to keep. Several modules in one act are judged together, so
+	// holders that depend on each other go on in one command.
+	shelf, before, err := seatDependenciesOnAssign(ctx, open, node, modules)
+	if err != nil {
+		return "", err
+	}
+	// **Before the new assignment can unsettle a seat somebody holds only by being alone**
+	// (novox/hq 04-ISSUES/170): what the mesh derived so far is written down, and then the
+	// assignment resolves against a record rather than against a coincidence.
+	settled, err := recordDerivedHolders(ctx, open)
+	if err != nil {
+		return "", err
+	}
+	var lines []string
+	var added []string
+	for _, module := range modules {
+		fresh, err := open.inventory.Assign(ctx, node, module)
+		if err != nil {
+			return strings.Join(lines, "\n"), err
+		}
+		if !fresh {
+			// Nothing changed, and saying "is assigned" would read as an action. One node runs one
+			// of each — the module's name is the assignment's identity (novox/hq ADR 0115).
+			lines = append(lines, fmt.Sprintf(
+				"%s already runs %s — one node runs one of each (ADR 0115); nothing changed", node, module))
+			continue
+		}
+		added = append(added, module)
+		lines = append(lines, fmt.Sprintf("%s is assigned %s", node, module))
+	}
+	if len(added) == 0 {
+		return strings.Join(lines, "\n"), nil
+	}
+	answer := strings.Join(lines, "\n")
+	for _, line := range settled {
+		answer += "\n  " + line
+	}
+	// What this act changed about this node's unmet seat dependencies, and nothing else (novox/hq
+	// ADR 0207): a dependency of a module just assigned, or one this assignment met. The rest of the
+	// node's list, and every other node's, is `status`'s.
+	for _, line := range unheldChange(shelf, node, before, append(append([]string(nil), before...), added...)) {
+		answer += "\n  " + line
+	}
+	// Its bus credential, in the same act (novox/hq issue 203): an assignment pushed before its
+	// credential exists delivers a process that cannot authenticate and crash-loops until somebody
+	// runs a second verb and a second push. Issued here when the module speaks on the bus and has
+	// no credential yet; kept when it has one, so re-assigning rotates nothing.
+	for _, module := range added {
+		if line := issueOnAssign(ctx, open, node, module); line != "" {
+			answer += "\n  " + line
+		}
+	}
+	plan, _, err := planFor(ctx, open, node)
+	if err != nil {
+		// Kept, and still refused. Both halves are the answer, and the rest of the mesh is still
+		// worth reporting: this machine's refusal is rarely the only consequence.
+		return answer + blockedElsewhere(ctx, open, node), err
+	}
+	// Kept, and cannot be hosted here. Said at once rather than discovered at push: a module whose
+	// capability the machine lacks is on the wrong machine, and the assignment records what a person
+	// meant while this line says it will not run until it moves. The rest of the node still pushes.
+	isAdded := map[string]bool{}
+	for _, m := range added {
+		isAdded[m] = true
+	}
+	for _, u := range plan.Unhostable {
+		if !isAdded[u.Module] {
+			continue
+		}
+		for _, c := range u.Missing {
+			answer += "\n  but " + catalogue.WrongMachine(u.Module, c, node)
+		}
+	}
+	return answer + fmt.Sprintf("\n  run `push %s` to send it", node) +
+		blockedElsewhere(ctx, open, node), nil
+}
+
+// seatDependenciesOnAssign is the refusal ADR 0207 makes at assignment, or nothing, with the
+// catalogue and the node's assignments it was judged against. Modules already assigned are not new
+// and are not judged again.
+func seatDependenciesOnAssign(ctx context.Context, open *stores, node string, modules []string) (
+	map[string]catalogue.Manifest, []string, error) {
+	shelf, err := open.inventory.Catalogue(ctx)
+	if err != nil {
+		return nil, nil, err
+	}
+	assigned, err := open.inventory.Assigned(ctx, node)
+	if err != nil {
+		return nil, nil, err
+	}
+	already := map[string]bool{}
+	for _, a := range assigned {
+		already[a] = true
+	}
+	var adding []string
+	for _, m := range modules {
+		if !already[m] {
+			adding = append(adding, m)
+		}
+	}
+	// The lines AssignRefusal says beside an assignment it lets through are said by unheldChange
+	// with everything else this act changed, so they are not said twice.
+	if _, err := catalogue.AssignRefusal(shelf, node, assigned, adding); err != nil {
+		return nil, nil, err
+	}
+	// Two modules declaring one package, path or unit is refused before anything is recorded
+	// (novox/hq ADR 0210, 04-ISSUES/235): kept, the node would not resolve until one came off again.
+	if err := catalogue.CollisionRefusal(shelf, node, assigned, adding); err != nil {
+		return nil, nil, err
+	}
+	return shelf, assigned, nil
+}
+
+// unassign takes modules off a node. What they leave behind is the host's business: a directory
+// holding anything the mesh did not put there is kept (novox/hq ADR 0030).
+//
+// It reports the rest of the mesh for the same reason assign does, and more sharply: taking a
+// module off one machine is the ordinary way to stop providing something to another, and nothing
+// about the command's own output would ever have said so.
+//
+// **Refused when it takes away the last holder of a seat a module left on the node depends on**
+// (novox/hq ADR 0207) — the other side of refusing that module's assignment without one. Several
+// modules in one act are judged together, so a holder and its dependents come off in one command.
+func unassign(ctx context.Context, open *stores, node string, modules ...string) (string, error) {
+	if len(modules) == 0 {
+		return "", fmt.Errorf("unassign %s names no module", node)
+	}
+	ctx, release, err := holdNodes(ctx, open, []string{node})
+	if err != nil {
+		return "", err
+	}
+	defer release()
+	shelf, err := open.inventory.Catalogue(ctx)
+	if err != nil {
+		return "", err
+	}
+	assigned, err := open.inventory.Assigned(ctx, node)
+	if err != nil {
+		return "", err
+	}
+	// Every one checked before any is taken off, so a refusal leaves the node as it was.
+	runs := map[string]bool{}
+	for _, a := range assigned {
+		runs[a] = true
+	}
+	for _, module := range modules {
+		if !runs[module] {
+			return "", fmt.Errorf("%s is not assigned to %s", module, node)
+		}
+	}
+	if err := catalogue.UnassignRefusal(shelf, node, assigned, modules); err != nil {
+		return "", err
+	}
+	for _, module := range modules {
+		if err := open.inventory.Unassign(ctx, node, module); err != nil {
+			return "", err
+		}
+	}
+	answer := fmt.Sprintf("%s no longer runs %s — run `push %s` to make it so",
+		node, strings.Join(modules, ", "), node)
+	var left []string
+	for _, a := range assigned {
+		if !slices.Contains(modules, a) {
+			left = append(left, a)
+		}
+	}
+	for _, line := range unheldChange(shelf, node, assigned, left) {
+		answer += "\n  " + line
+	}
+	return answer + blockedElsewhere(ctx, open, node), nil
+}
+
+// splitModules is a surface's one `module` field as the modules it names: several, comma-separated,
+// are one act (novox/hq ADR 0207), so the holders that depend on each other go on together from the
+// command API and the controller seat's verbs as they do from the command line.
+func splitModules(field string) []string {
+	var out []string
+	for _, m := range strings.Split(field, ",") {
+		if m = strings.TrimSpace(m); m != "" {
+			out = append(out, m)
+		}
+	}
+	return out
+}
+
+// blockedElsewhere is every OTHER machine that cannot be worked out as things now stand.
+//
+// **The state, not the cause.** Saying "this assignment broke laptop" would mean resolving the
+// whole mesh twice and would still be a guess about which of several changes did it; saying
+// "laptop cannot be worked out, and here is what it says" is true, is what somebody has to fix,
+// and cannot mislead. The machine that was just changed is left out because its own refusal is
+// already the answer beside this one.
+//
+// Nothing here can fail the act it reports on. A mesh that cannot be read is worth saying and is
+// not a reason to claim the assignment did not happen — it did.
+func blockedElsewhere(ctx context.Context, open *stores, except string) string {
+	nodes, err := open.inventory.Nodes(ctx)
+	if err != nil {
+		return "\n\nThe rest of the mesh could not be checked: " + err.Error()
+	}
+	blocked := map[string]string{}
+	for _, n := range nodes {
+		if n.Name == except {
+			continue
+		}
+		if _, _, err := planFor(ctx, open, n.Name); err != nil {
+			blocked[n.Name] = err.Error()
+		}
+	}
+	if len(blocked) == 0 {
+		return ""
+	}
+	names := make([]string, 0, len(blocked))
+	for name := range blocked {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+
+	var out strings.Builder
+	fmt.Fprintf(&out, "\n\nAND %d other machine(s) cannot be worked out as things stand, so "+
+		"nothing will be sent to them:\n", len(names))
+	for _, name := range names {
+		fmt.Fprintf(&out, "  %s\n", name)
+		for _, line := range strings.Split(strings.TrimRight(blocked[name], "\n"), "\n") {
+			fmt.Fprintf(&out, "    %s\n", strings.TrimSpace(line))
+		}
+	}
+	out.WriteString("\nThis may or may not be what just changed — it is what is true now.")
+	return out.String()
+}
+
+// issueOnAssign gives a newly assigned module its bus credential, the way `module issue` does, and
+// says what it did in one line. Nothing for a module that declares no broker secret; nothing for one
+// whose user is already minted (a credential is rotated on purpose, never by re-assigning); and when
+// the bus cannot be reached from here, the line names the verb and the push that would refuse the
+// module until it is run — never a silent placeholder (novox/hq issue 203).
+func issueOnAssign(ctx context.Context, open *stores, node, module string) string {
+	inv := open.inventory
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return ""
+	}
+	m, known := shelf[module]
+	if !known || mayIssue(m) != nil {
+		return ""
+	}
+	user := broker.Principal{Kind: broker.KindModule, Node: node, Module: module}.Username()
+	if _, minted, err := inv.BusUserHash(ctx, user); err != nil || minted {
+		return ""
+	}
+	busAddress, err := broker.BusAddress()
+	if err == nil {
+		err = issueOnTheNewBus(ctx, inv, m, node, busAddress)
+	}
+	if err != nil {
+		return fmt.Sprintf("its bus credential is not issued (%v): `module issue %s --node %s` first — "+
+			"`push %s` refuses to send %s until it is", err, module, node, node, module)
+	}
+	return fmt.Sprintf("its bus credential is issued and sealed to %s, and arrives with the push", node)
+}

cmd/mesh-controller/acts_test.go

@@ -0,0 +1,133 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// What an assignment says about the machines it was not about.
+
+// **An assignment that blocks another machine says so.**
+//
+// The shape found on a four-node raise: a module offering a mesh-scoped provision stops offering it
+// the moment its own node stops resolving, so an assignment to the provider's machine silently took
+// a provision away from every consumer elsewhere. Those consumers were then told *nothing in this
+// mesh provides it*, naming as the remedy a module that was already assigned — which read, on the
+// way back, as a version bump breaking provider recognition. It was neither the version nor the
+// provider: it was one machine's set of assignments, and nothing said so.
+//
+// novox/hq 04-ISSUES/017 names the general shape — an action succeeding into a state its own verify
+// rejects, because the action's test is not the test the verify uses. `assign` tested one node; the
+// verify is resolution over all of them.
+func TestAnAssignmentThatBlocksAnotherMachineSaysWhichAndWhy(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+
+	// A provider on the hub and a consumer on the other machine, both resolving.
+	register(t, open, catalogue.Manifest{Module: "step-ca", Version: "1",
+		Provides: []catalogue.Offer{{Name: "acme-ca", Scope: catalogue.ScopeMesh}},
+		Serves:   map[string]map[string]any{"acme-ca": {"path": "/acme/directory"}}})
+	register(t, open, catalogue.Manifest{Module: "route-proxy", Version: "1",
+		Requires: []string{"acme-ca"}})
+	if _, err := assign(ctx, open, "anchor", "step-ca"); err != nil {
+		t.Fatal(err)
+	}
+	said, err := assign(ctx, open, "laptop", "route-proxy")
+	if err != nil {
+		t.Fatalf("a mesh that should resolve did not: %v\n%s", err, said)
+	}
+	if strings.Contains(said, "cannot be worked out") {
+		t.Fatalf("a well mesh was reported as blocked:\n%s", said)
+	}
+
+	// Now break the hub's own set, with nothing but the command a person has.
+	one, two := rivals()
+	register(t, open, one)
+	register(t, open, two)
+	if _, err := assign(ctx, open, "anchor", "rival-one"); err != nil {
+		t.Fatal(err)
+	}
+	said, err = assign(ctx, open, "anchor", "rival-two")
+	if err == nil {
+		t.Fatal("an assignment that makes its own node incoherent was not reported at all")
+	}
+
+	// The node's own refusal is the error, as it always was. What is new is that the machines this
+	// just took a provision away from are named in the same breath.
+	if !strings.Contains(said, "laptop") {
+		t.Fatalf("the machine this blocked is not named:\n%s\n\n%v", said, err)
+	}
+	if !strings.Contains(said, "acme-ca") {
+		t.Fatalf("what laptop is now missing is not said:\n%s", said)
+	}
+	if !strings.Contains(said, "cannot be worked out as things stand") {
+		t.Fatalf("the report does not say what state the mesh is in:\n%s", said)
+	}
+	// And the assignment is kept: it is what a person meant, and assignment is not an ordering.
+	assigned, err := open.inventory.Assigned(ctx, "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !contains(assigned, "rival-two") {
+		t.Fatalf("the assignment was not kept: %v", assigned)
+	}
+}
+
+// A consumer assigned before its provider is refused for itself and kept, because assignment is not
+// an ordering — refusing the first half of a pair would make the order somebody types two commands
+// in part of the mesh's rules.
+func TestAConsumerAssignedBeforeItsProviderIsStillAssigned(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, catalogue.Manifest{Module: "route-proxy", Version: "1",
+		Requires: []string{"acme-ca"}})
+
+	said, err := assign(ctx, open, "laptop", "route-proxy")
+	if err == nil {
+		t.Fatalf("a consumer with nothing to consume resolved:\n%s", said)
+	}
+	assigned, err := open.inventory.Assigned(ctx, "laptop")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !contains(assigned, "route-proxy") {
+		t.Fatalf("assignment became an ordering: %v", assigned)
+	}
+}
+
+// Unassigning is the ordinary way to stop providing something to another machine, and it reports
+// the same way for the same reason.
+func TestUnassigningAProviderNamesWhoIsNowBlocked(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, catalogue.Manifest{Module: "step-ca", Version: "1",
+		Provides: []catalogue.Offer{{Name: "acme-ca", Scope: catalogue.ScopeMesh}},
+		Serves:   map[string]map[string]any{"acme-ca": {"path": "/acme/directory"}}})
+	register(t, open, catalogue.Manifest{Module: "route-proxy", Version: "1",
+		Requires: []string{"acme-ca"}})
+	if _, err := assign(ctx, open, "anchor", "step-ca"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := assign(ctx, open, "laptop", "route-proxy"); err != nil {
+		t.Fatal(err)
+	}
+
+	said, err := unassign(ctx, open, "anchor", "step-ca")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(said, "laptop") || !strings.Contains(said, "acme-ca") {
+		t.Fatalf("taking the provider away said nothing about who was consuming it:\n%s", said)
+	}
+}
+
+func contains(all []string, one string) bool {
+	for _, s := range all {
+		if s == one {
+			return true
+		}
+	}
+	return false
+}

cmd/mesh-controller/addresses_test.go

@@ -0,0 +1,352 @@
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/licences"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// An address is read from the node's settings where it is used, never recorded with a port
+// (novox/hq 04-ISSUES/102). Three readers did not follow the setting; each is held to it here.
+
+var aDigest = "sha256:" + strings.Repeat("e", 64)
+
+// **A build is recorded by digest and path**, whatever address the builder pushed to — what it
+// made and what it stood on both; an image the module runs from elsewhere is left where it says.
+func TestABuildIsRecordedWithoutTheStoresAddress(t *testing.T) {
+	manifest, _ := json.Marshal(map[string]any{
+		"module": "gitea", "version": "1",
+		"resources": []map[string]any{
+			{"id": "server", "type": "container", "name": "mesh-gitea",
+				"image": "anchor.internal:5100/gitea/server@" + aDigest},
+			{"id": "config", "type": "archive", "path": "/etc/gitea", "digest": aDigest,
+				"source": "http://anchor.internal:5100/v2/gitea/config/blobs/" + aDigest},
+			{"id": "cache", "type": "container", "name": "mesh-gitea-cache",
+				"image": "valkey/valkey@" + aDigest},
+		},
+	})
+	kept := buildFrom(link.BuildResult{
+		ID: "b1", Repository: "https://forge.example/gitea.git", Commit: "abc", On: "laptop",
+		Manifest: manifest,
+		Made: []link.MadeArtifact{
+			{Name: "server", Kind: "image", Reference: "anchor.internal:5100/gitea/server@" + aDigest},
+			{Name: "config", Kind: "archive", Reference: "http://anchor.internal:5100/v2/gitea/config/blobs/" + aDigest},
+		},
+		Against: []string{"anchor.internal:5100/mesh-tools/runtime@" + aDigest},
+	})
+	if kept.Module != "gitea" {
+		t.Fatalf("the module was not read from the recorded manifest: %q", kept.Module)
+	}
+	if kept.Made[0].Reference != catalogue.ArtifactStoreScheme+"gitea/server@"+aDigest {
+		t.Errorf("the image is recorded as %q, address and all", kept.Made[0].Reference)
+	}
+	if kept.Made[1].Reference != catalogue.ArtifactStoreScheme+"gitea/config/blobs/"+aDigest {
+		t.Errorf("the archive is recorded as %q, address and all", kept.Made[1].Reference)
+	}
+	recorded, err := catalogue.ParseManifest(kept.Manifest)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := recorded.Resources[0]["image"]; got != catalogue.ArtifactStoreScheme+"gitea/server@"+aDigest {
+		t.Errorf("the recorded manifest's image is %v", got)
+	}
+	if got := recorded.Resources[1]["source"]; got != catalogue.ArtifactStoreScheme+"gitea/config/blobs/"+aDigest {
+		t.Errorf("the recorded manifest's archive is %v", got)
+	}
+	if got := recorded.Resources[2]["image"]; got != "valkey/valkey@"+aDigest {
+		t.Errorf("an image the build did not make was rewritten: %v", got)
+	}
+	if strings.Contains(string(kept.Manifest), "anchor.internal:5100") {
+		t.Errorf("the recorded manifest still carries the store's address:\n%s", kept.Manifest)
+	}
+	// What the build stood on is an edge to another module's artifact, and it is recorded the way
+	// that artifact is: by path in the store, so the edge still names the same thing when the
+	// store answers at another address.
+	if kept.Against[0] != catalogue.ArtifactStoreScheme+"mesh-tools/runtime@"+aDigest {
+		t.Errorf("what the build stood on was recorded by address: %v", kept.Against)
+	}
+}
+
+// aStore is a module offering the artifact store on 5000, published the long way as the
+// distribution module does, so a node may be given another number for it.
+func aStore() catalogue.Manifest {
+	return catalogue.Manifest{Module: "distribution", Version: "1",
+		Provides: []catalogue.Offer{{Name: catalogue.ArtifactStoreProvision, Scope: catalogue.ScopeMesh}},
+		Serves:   map[string]map[string]any{catalogue.ArtifactStoreProvision: {"port": float64(5000)}},
+		Listens:  []catalogue.Listening{{Port: 5000, From: catalogue.FromMesh}},
+		Resources: []map[string]any{{"id": "store", "type": "container", "name": "mesh-registry",
+			"ports": []any{"5000:5000"}, "image": "registry@" + aDigest}}}
+}
+
+// **The trust a machine writes for the store, and the address every built image is fetched
+// through, say the port the node gave the store** — not the catalogue's number.
+func TestTheRegistryTrustAndEveryImageFollowThePortTheNodeGaveTheStore(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, aStore())
+	if _, err := assign(ctx, open, "anchor", "distribution"); err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.SetSettings(ctx, "anchor", "distribution",
+		map[string]any{catalogue.PortsSetting: map[string]any{"5000": 5101}}); err != nil {
+		t.Fatal(err)
+	}
+	// A module the mesh built, recorded by digest and path, running on the other machine.
+	if err := open.inventory.RecordBuild(ctx, inventory.Build{
+		ID: "b1", Repository: "r", Module: "app", Commit: "abc",
+		Made: []inventory.Artifact{{Name: "server", Kind: "image",
+			Reference: catalogue.ArtifactStoreScheme + "app/server@" + aDigest}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	register(t, open, catalogue.Manifest{Module: "app", Version: "1",
+		Resources: []map[string]any{{"id": "server", "type": "container", "name": "mesh-app",
+			"image": catalogue.ArtifactStoreScheme + "app/server@" + aDigest}}})
+	if _, err := assign(ctx, open, "laptop", "app"); err != nil {
+		t.Fatal(err)
+	}
+	// The runtime's trust is the runtime's module's to write (novox/hq ADR 0222): a stand-in for it
+	// asks where this machine reaches the store, as the docker module does.
+	register(t, open, catalogue.Manifest{Module: "runtime", Version: "1",
+		Resources: []map[string]any{{"id": "daemon", "type": "file", "path": "/etc/docker/daemon.json",
+			"into": "json", "content": `{"insecure-registries": ["${seat:mesh-artifact-store:reach}"]}` + "\n"}}})
+	if _, err := assign(ctx, open, "laptop", "runtime"); err != nil {
+		t.Fatal(err)
+	}
+
+	on := map[string]bool{"anchor": true, "laptop": true}
+	node, port, found, err := artifactStoreOnNetwork(ctx, open.inventory, on)
+	if err != nil || !found || node != "anchor" || port != "5101" {
+		t.Fatalf("the store is found on %q:%q (%v, %v); the node put it on 5101", node, port, found, err)
+	}
+
+	var trust string
+	for _, r := range composed(t, open, "laptop").Resources {
+		if r["path"] == "/etc/docker/daemon.json" {
+			trust, _ = r["content"].(string)
+		}
+		if r["id"] == "app.server" && r["image"] != "anchor.internal:5101/app/server@"+aDigest {
+			t.Errorf("the image the mesh built is fetched as %v", r["image"])
+		}
+	}
+	if !strings.Contains(trust, "anchor.internal:5101") || strings.Contains(trust, ":5000") {
+		t.Fatalf("the runtime is told to trust %q; the node put the store on 5101", trust)
+	}
+
+	// And a replay to the catalogue says where the store is now.
+	announced, err := following{open}.Announceable(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(announced) != 1 || announced[0].Made[0].Reference != "anchor.internal:5101/app/server@"+aDigest {
+		t.Fatalf("the replay announces %+v", announced)
+	}
+}
+
+// **The control plane's own connections say the port the node gave the store and the broker.**
+//
+// Composed from the control plane's own manifest against a real inventory: the store's module is
+// given 6852 on this node the way genesis or an operator gives it, and the control plane's
+// process is told so beside the sealed connection genesis wrote.
+func TestTheControlPlaneIsToldWhereTheNodePutTheStoreAndTheBroker(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	raw, err := os.ReadFile("../../module.json")
+	if err != nil {
+		t.Fatal(err)
+	}
+	m, err := catalogue.ParseManifest(raw)
+	if err != nil {
+		t.Fatal(err)
+	}
+	control, err := withSeatPorts(m).Resolve([]catalogue.Built{{Name: "controller", Kind: catalogue.ArtifactBundle,
+		Reference: "https://registry.example/mesh-controller/controller.tar.gz", Digest: aDigest}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	register(t, open, control)
+	register(t, open, catalogue.Manifest{Module: "postgres", Version: "1",
+		Claims:  []catalogue.Claim{{Name: "mesh-store", Scope: catalogue.ScopeMesh}},
+		Listens: []catalogue.Listening{{Port: 5432, From: catalogue.FromMesh}},
+		Resources: []map[string]any{{"id": "server", "type": "container", "name": "mesh-store",
+			"ports": []any{"5432:5432"}, "image": "pg@" + aDigest}}})
+	register(t, open, catalogue.Manifest{Module: "lavinmq", Version: "1",
+		Claims: []catalogue.Claim{{Name: "mesh-broker", Scope: catalogue.ScopeMesh}},
+		Listens: []catalogue.Listening{{Port: 5671, From: catalogue.FromMesh},
+			{Port: 5672, From: catalogue.FromMesh}},
+		Guards: []int{15672},
+		Resources: []map[string]any{{"id": "server", "type": "container", "name": "mesh-broker",
+			"ports": []any{"5671:5671", "5672:5672", "127.0.0.1:15672:15672"}, "image": "mq@" + aDigest}}})
+	// The control plane's own bus user is the installer's, seeded at genesis before the controller
+	// runs (SeedBusUser); without it a push now refuses the credential nobody issued (issue 203).
+	if err := open.inventory.SeedBusUser(ctx, inventory.BusUser{Username: "anchor.mesh-controller",
+		Kind: inventory.BusController, Node: "anchor", Module: "mesh-controller"}, "bootstrap"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := assign(ctx, open, "anchor", "mesh-controller"); err != nil {
+		t.Fatal(err)
+	}
+	// The store's module is registered and given its port, and NOT assigned: the state genesis
+	// leaves a given-port node in before the foundation is adopted as modules (04-ISSUES/085).
+	if err := open.inventory.SetSettings(ctx, "anchor", "postgres",
+		map[string]any{catalogue.PortsSetting: map[string]any{"5432": 6852}}); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := assign(ctx, open, "anchor", "lavinmq"); err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.SetSettings(ctx, "anchor", "lavinmq",
+		map[string]any{catalogue.PortsSetting: map[string]any{"5672": 5679}}); err != nil {
+		t.Fatal(err)
+	}
+
+	var env map[string]any
+	for _, r := range composed(t, open, "anchor").Resources {
+		if r["id"] == "mesh-controller.controller" {
+			env, _ = r["env"].(map[string]any)
+		}
+	}
+	if env == nil {
+		t.Fatal("the control plane's process is not in its own node's declaration")
+	}
+	for key, want := range map[string]string{
+		"MESH_STORE_INVENTORY_PORT": "6852",
+		"MESH_STORE_IDENTITY_PORT":  "6852",
+		"MESH_STORE_LICENCES_PORT":  "6852",
+		"MESH_BROKER_AMQP_PORT":     "5679",
+		// Neither given nor assigned by the mesh: the manifest's own number is NOT the answer,
+		// because the sealed value beside it carries the port genesis wrote (finding F3).
+		"MESH_BROKER_ADDRESS_PORT":    "",
+		"MESH_BROKER_MANAGEMENT_PORT": "",
+	} {
+		if env[key] != want {
+			t.Errorf("the control plane is told %s=%v; the node says %q", key, env[key], want)
+		}
+	}
+}
+
+// withSeatPorts is the control plane's manifest with the seat placeholders in its environment —
+// added here until module.json carries them (the manifest lands one commit after the code that
+// fills it, so a control plane one build behind never sees a placeholder it cannot fill).
+func withSeatPorts(m catalogue.Manifest) catalogue.Manifest {
+	seatPorts := map[string]string{
+		"MESH_STORE_INVENTORY_PORT":   "${seat:mesh-store:5432}",
+		"MESH_STORE_IDENTITY_PORT":    "${seat:mesh-store:5432}",
+		"MESH_STORE_LICENCES_PORT":    "${seat:mesh-store:5432}",
+		"MESH_BROKER_AMQP_PORT":       "${seat:mesh-broker:5672}",
+		"MESH_BROKER_MANAGEMENT_PORT": "${seat:mesh-broker:15672}",
+		"MESH_BROKER_ADDRESS_PORT":    "${seat:mesh-broker:5671}",
+	}
+	out := m
+	out.Resources = nil
+	for _, r := range m.Resources {
+		if r["type"] != "container" && r["type"] != "process" {
+			out.Resources = append(out.Resources, r)
+			continue
+		}
+		copied := map[string]any{}
+		for k, v := range r {
+			copied[k] = v
+		}
+		env := map[string]any{}
+		if had, ok := r["env"].(map[string]any); ok {
+			for k, v := range had {
+				env[k] = v
+			}
+		}
+		for k, v := range seatPorts {
+			if _, said := env[k]; !said {
+				env[k] = v
+			}
+		}
+		copied["env"] = env
+		out.Resources = append(out.Resources, copied)
+	}
+	return out
+}
+
+// aLoneNode is one capable machine with nothing placed on any network — the control-node during
+// genesis, before the "network" step, which is after the store, the broker, the vault and the
+// catalogue have each been built and pushed (finding F2).
+func aLoneNode(t *testing.T) *stores {
+	t.Helper()
+	inventory.ForTest(t)
+	licences.ForTest(t)
+	open, err := openStores(t.Context())
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(open.Close)
+	for _, m := range provided {
+		if err := open.inventory.Provide(t.Context(), m); err != nil {
+			t.Fatal(err)
+		}
+	}
+	record, err := open.inventory.AddNode(t.Context(), "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	reported, _ := json.Marshal(map[string]any{"capabilities": []map[string]any{
+		{"name": "container-runtime", "present": true}}})
+	var profile map[string]any
+	_ = json.Unmarshal(reported, &profile)
+	if err := open.inventory.RecordProfile(t.Context(), record.ID, profile); err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.RecordSealingKey(t.Context(), record.ID, aPublicKey(t)); err != nil {
+		t.Fatal(err)
+	}
+	return open
+}
+
+// **Before the network exists, the store's own node reaches it by loopback** — never refused,
+// never handed the scheme: a genesis pushes the store, the broker, the vault and the catalogue to
+// a node on no network, and builds the catalogue on a base it must be able to pull.
+func TestOnANodeWithNoNetworkTheStoreIsReachedByLoopback(t *testing.T) {
+	open := aLoneNode(t)
+	ctx := t.Context()
+	register(t, open, aStore())
+	register(t, open, catalogue.Manifest{Module: "builder", Version: "1",
+		Requires: []string{catalogue.ArtifactStoreProvision},
+		Resources: []map[string]any{{"id": "server", "type": "container", "name": "mesh-builder",
+			"image": "registry.example/mesh-builder@" + aDigest}}})
+	if err := open.inventory.RecordBuild(ctx, inventory.Build{
+		ID: "b1", Repository: "r", Module: "postgres", Commit: "abc",
+		Made: []inventory.Artifact{{Name: "runtime", Kind: "image",
+			Reference: catalogue.ArtifactStoreScheme + "postgres/runtime@" + aDigest}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	register(t, open, catalogue.Manifest{Module: "postgres", Version: "1",
+		Resources: []map[string]any{{"id": "runtime", "type": "container", "name": "mesh-postgres",
+			"image": catalogue.ArtifactStoreScheme + "postgres/runtime@" + aDigest}}})
+	for _, module := range []string{"distribution", "builder", "postgres"} {
+		if _, err := assign(ctx, open, "anchor", module); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := open.inventory.SetSettings(ctx, "anchor", "distribution",
+		map[string]any{catalogue.PortsSetting: map[string]any{"5000": 5100}}); err != nil {
+		t.Fatal(err)
+	}
+
+	var image any
+	for _, r := range composed(t, open, "anchor").Resources {
+		if r["id"] == "postgres.runtime" {
+			image = r["image"]
+		}
+	}
+	if image != "127.0.0.1:5100/postgres/runtime@"+aDigest {
+		t.Fatalf("on the store's own node, off any network, the image is fetched as %v", image)
+	}
+	held := heldBy(ctx)
+	if got := held["postgres/runtime"]; got != "127.0.0.1:5100/postgres/runtime@"+aDigest {
+		t.Fatalf("a builder beside the store is handed the base %q", got)
+	}
+}

cmd/mesh-controller/adopting_test.go

@@ -0,0 +1,641 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+	"github.com/novox/mesh-controller/internal/overlay"
+)
+
+// novox/hq ADR 0100: taking a module is its cutover; converging a node is one act, previewed, and
+// refused while a found container is held; returning to adopted keeps what was taken.
+
+// anAdoptedAnchor is aMesh with the anchor adopted, running a served module the predecessor also
+// runs and a module with only a file, and a filter module in the catalogue.
+func anAdoptedAnchor(t *testing.T) (*stores, *[]string) {
+	t.Helper()
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, catalogue.Manifest{Module: "hello-web", Version: "1",
+		Listens: []catalogue.Listening{{Port: 8080, From: catalogue.FromEverywhere}},
+		Resources: []map[string]any{
+			{"id": "page", "type": "file", "path": "/var/lib/hello-web/index.html", "content": "hi"},
+			{"id": "server", "type": "container", "name": "hello-web", "ports": []any{"8080:80"},
+				"image": "registry.example/hello@sha256:" + strings.Repeat("a", 64)},
+		}})
+	register(t, open, catalogue.Manifest{Module: "notes", Version: "1",
+		Resources: []map[string]any{
+			{"id": "conf", "type": "file", "path": "/etc/notes.conf", "content": "x"},
+		}})
+	register(t, open, catalogue.Manifest{Module: "nftables", Version: "1",
+		Filtering: &catalogue.Filtering{Into: "/etc/nftables.conf"},
+		Resources: []map[string]any{{"id": "load", "type": "service", "unit": "mesh-filter.service",
+			"state": "running", "restart-on": []any{"filtering"}}}})
+	if err := open.inventory.SetAdopted(ctx, "anchor", true); err != nil {
+		t.Fatal(err)
+	}
+	for _, m := range []string{"hello-web", "notes"} {
+		if _, err := assign(ctx, open, "anchor", m); err != nil {
+			t.Fatal(err)
+		}
+	}
+	sent := &[]string{}
+	saved := sendNodes
+	sendNodes = func(_ context.Context, _ *stores, names []string) error {
+		*sent = append(*sent, names...)
+		return nil
+	}
+	t.Cleanup(func() { sendNodes = saved })
+	return open, sent
+}
+
+// reportsHolding has the anchor report, on what it was last sent, holding what is given.
+func reportsHolding(t *testing.T, open *stores, held ...link.Held) {
+	t.Helper()
+	reportsReaching(t, open, []link.Reach{
+		{Protocol: "tcp", Address: "0.0.0.0", Port: 22, By: "sshd"},
+		{Protocol: "tcp", Address: "0.0.0.0", Port: 8080, By: "hello-web", Published: true,
+			ContainerPort: 80},
+		{Protocol: "tcp", Address: "0.0.0.0", Port: 5000, By: "predecessor-registry",
+			Published: true, ContainerPort: 5000},
+		{Protocol: "tcp", Address: "127.0.0.1", Port: 15672, By: "mesh-broker",
+			Published: true, ContainerPort: 15672},
+	}, held...)
+}
+
+// reportsReaching has the anchor report, on what it was last sent, what is reachable on it and
+// holding what is given.
+func reportsReaching(t *testing.T, open *stores, reachable []link.Reach, held ...link.Held) {
+	t.Helper()
+	ctx := t.Context()
+	body, err := composed(t, open, "anchor").Body()
+	if err != nil {
+		t.Fatal(err)
+	}
+	record, err := open.inventory.NodeByName(ctx, "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.RecordSent(ctx, record.ID, digestOf(body), nil); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := (link.Enrolment{Inventory: open.inventory}).Heard(ctx, link.Report{
+		Node: "anchor", Applied: []string{"hello-web.x"}, Declared: digestOf(body),
+		Firewall: "ufw", Held: held, Reachable: reachable,
+		// A machine says which of its links face outside on every apply (novox/hq ADR 0140), and a
+		// filter is not sent to one that has not. The anchor reports one, as a real host does; this
+		// fixture lacked it from 2026-09-28 and nothing ran the test (issue 177).
+		Outward: []string{"eth0"},
+		// And what filters it (ADR 0168): its front end, the runtime's own, and a chain a
+		// predecessor left in the runtime's user chain.
+		Filters: anchorFilters,
+	}); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// anchorFilters is what the adopted anchor says filters it: ufw's chains, the runtime's, and a
+// predecessor's chain the mesh did not write.
+var anchorFilters = []link.Filter{
+	{Where: "table ip filter, chain ufw-reject-input", Owner: "found-firewall", Refuses: "reject"},
+	{Where: "table ip filter, chain DOCKER", Owner: "runtime", Refuses: `iifname != "docker0" oifname "docker0" drop`},
+	{Where: "table ip filter, chain DOCKER-USER", Owner: "other", Refuses: `iifname "eth0" tcp dport 6000 drop`},
+}
+
+var (
+	heldContainer = link.Held{ID: "hello-web.server", Module: "hello-web", Kind: "container",
+		Target: "hello-web", Since: time.Now()}
+	heldFile = link.Held{ID: "notes.conf", Module: "notes", Kind: "file",
+		Target: "/etc/notes.conf", Since: time.Now(), Kept: "/var/lib/mesh-host/kept/abc-notes.conf"}
+)
+
+func TestTakingAModuleNotOnTheNodeIsRefused(t *testing.T) {
+	open, _ := anAdoptedAnchor(t)
+	if _, err := take(t.Context(), open, "anchor", "nftables", takeOptions{Yes: true}); !errors.Is(err, inventory.ErrNotAssigned) {
+		t.Fatalf("taking an unassigned module gave %v", err)
+	}
+	if _, err := take(t.Context(), open, "laptop", "network", takeOptions{Yes: true}); !errors.Is(err, inventory.ErrNotAdopted) {
+		t.Fatalf("taking on a converged node gave %v", err)
+	}
+}
+
+func TestConvergingIsRefusedOnAPreviewThatWouldBeStale(t *testing.T) {
+	open, sent := anAdoptedAnchor(t)
+	_, err := converge(t.Context(), open, "anchor", false, "", "")
+	if err == nil || !strings.Contains(err.Error(), "has not reported") {
+		t.Fatalf("a node that never reported was previewed: %v", err)
+	}
+	if len(*sent) != 0 {
+		t.Fatal("a refused converge sent something")
+	}
+}
+
+func TestTheFlipIsRefusedWhileAFoundContainerIsHeld(t *testing.T) {
+	open, sent := anAdoptedAnchor(t)
+	reportsHolding(t, open, heldContainer, heldFile)
+	_, err := converge(t.Context(), open, "anchor", true, "", "")
+	if err == nil || !strings.Contains(err.Error(), "take anchor hello-web once its data has moved") {
+		t.Fatalf("the flip was not refused while hello-web holds its found container: %v", err)
+	}
+	if n, _ := open.inventory.NodeByName(t.Context(), "anchor"); !n.Adopted || len(*sent) != 0 {
+		t.Fatal("a refused flip changed something")
+	}
+}
+
+func TestTakingNamesWhatItReplaces(t *testing.T) {
+	open, _ := anAdoptedAnchor(t)
+	reportsHolding(t, open, heldContainer, heldFile)
+	ctx := t.Context()
+	// The machine holds something for the module, so the take acts on the preview the operator
+	// saw and names its digest (novox/hq ADR 0163).
+	preview, err := take(ctx, open, "anchor", "hello-web", takeOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	saw := takeDigestIn(t, preview)
+	if !strings.Contains(preview, "nothing taken; `take anchor hello-web --yes "+saw+"`") {
+		t.Fatalf("the preview does not say how to act on it:\n%s", preview)
+	}
+	if taken, _ := open.inventory.Taken(ctx, "anchor"); len(taken) != 0 {
+		t.Fatal("the preview took something")
+	}
+	if _, err := take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true}); err == nil || !strings.Contains(err.Error(), "name its digest") {
+		t.Fatalf("--yes without the digest was not refused: %v", err)
+	}
+	said, err := take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true, Digest: saw})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(said, "container hello-web (hello-web.server)") ||
+		!strings.Contains(said, "push anchor") {
+		t.Fatalf("taking did not say what it replaces and what to run:\n%s", said)
+	}
+}
+
+// takeDigestIn is the digest a take's preview printed.
+func takeDigestIn(t *testing.T, preview string) string {
+	t.Helper()
+	for _, line := range strings.Split(preview, "\n") {
+		if fields := strings.Fields(line); len(fields) == 2 && fields[0] == "preview" {
+			return fields[1]
+		}
+	}
+	t.Fatalf("the preview printed no digest:\n%s", preview)
+	return ""
+}
+
+// A take acts on the preview the operator saw, and on an account of the machine that is still the
+// machine: a changed preview and a stale account refuse (novox/hq ADR 0163, rule 1).
+func TestATakeIsRefusedOnAChangedPreviewOrAStaleAccount(t *testing.T) {
+	open, _ := anAdoptedAnchor(t)
+	ctx := t.Context()
+	reportsHolding(t, open, heldContainer, heldFile)
+	preview, err := take(ctx, open, "anchor", "hello-web", takeOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	saw := takeDigestIn(t, preview)
+
+	// The machine reports again, and what it holds has changed: the found container now carries
+	// facts the preview never showed.
+	changed := heldContainer
+	changed.Facts = map[string]any{"image": "hello:2", "declared_image": "registry.example/hello"}
+	reportsHolding(t, open, changed, heldFile)
+	_, err = take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true, Digest: saw})
+	if err == nil || !strings.Contains(err.Error(), "has changed since preview "+saw) {
+		t.Fatalf("a changed preview was acted on: %v", err)
+	}
+	if taken, _ := open.inventory.Taken(ctx, "anchor"); len(taken) != 0 {
+		t.Fatal("a refused take took something")
+	}
+
+	// And an account older than the flip allows.
+	preview, err = take(ctx, open, "anchor", "hello-web", takeOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	saw = takeDigestIn(t, preview)
+	saved := reportFreshFor
+	reportFreshFor = -time.Second
+	defer func() { reportFreshFor = saved }()
+	_, err = take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true, Digest: saw})
+	if err == nil || !strings.Contains(err.Error(), "a take acts only on an account newer than") {
+		t.Fatalf("a stale account was acted on: %v", err)
+	}
+	reportFreshFor = saved
+	if _, err := take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true, Digest: saw}); err != nil {
+		t.Fatal(err)
+	}
+	// A module the machine holds nothing for has nothing to compare: --yes alone suffices.
+	if _, err := take(ctx, open, "anchor", "notes", takeOptions{Yes: true}); err == nil {
+		// notes holds a file, so this one needs the digest too.
+		t.Fatal("notes holds a found file and was taken without a digest")
+	}
+}
+
+func TestConvergingPreviewsThenChangesAndAdoptingKeepsWhatWasTaken(t *testing.T) {
+	open, sent := anAdoptedAnchor(t)
+	ctx := t.Context()
+	if _, err := take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true}); err != nil {
+		t.Fatal(err)
+	}
+	reportsHolding(t, open, heldFile)
+
+	preview, err := converge(ctx, open, "anchor", false, "", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, want := range []string{
+		"tcp/8080 hello-web (published, container port 80)",
+		"declared by hello-web (from anywhere)",
+		"WILL CLOSE — no module assigned here declares it",
+		// The anchor faces inward and is on the private network: the derived filter admits ssh
+		// from the mesh only.
+		"WILL CLOSE to everything outside the private network — ssh stays open from the mesh",
+		"notes\n      replacing the found file /etc/notes.conf (notes.conf), original kept at",
+		"assigns nftables",
+		"the found firewall (ufw) is disabled, never flushed",
+		// What it routes is not a listener: said not to be previewed, and to be dropped.
+		"not previewed: traffic the machine routes that is not a published port",
+		"the derived filter drops it unless a module declares it",
+	} {
+		if !strings.Contains(preview, want) {
+			t.Errorf("the preview does not say %q:\n%s", want, preview)
+		}
+	}
+	if strings.Contains(preview, "15672") {
+		t.Errorf("a loopback listener is in the preview:\n%s", preview)
+	}
+	// What filters the machine now, and the fate of each (novox/hq ADR 0168): the predecessor's
+	// chain is named as not the mesh's and left, so the reader knows before the flip.
+	for _, want := range []string{
+		"table ip filter, chain DOCKER-USER",
+		"NOT THE MESH'S; left in force",
+		`iifname "eth0" tcp dport 6000 drop`,
+		"table ip filter, chain ufw-reject-input",
+		"the found firewall's; retired with it",
+		"the container runtime's own; left",
+	} {
+		if !strings.Contains(preview, want) {
+			t.Errorf("the preview does not say %q:\n%s", want, preview)
+		}
+	}
+	for _, line := range strings.Split(preview, "\n") {
+		if strings.Contains(line, "5000") && !strings.Contains(line, "WILL CLOSE") {
+			t.Errorf("an undeclared published port is not said to close: %s", line)
+		}
+	}
+	if n, _ := open.inventory.NodeByName(ctx, "anchor"); !n.Adopted || len(*sent) != 0 {
+		t.Fatal("the preview changed something")
+	}
+
+	if _, err := converge(ctx, open, "anchor", true, digestIn(t, preview), ""); err != nil {
+		t.Fatal(err)
+	}
+	n, _ := open.inventory.NodeByName(ctx, "anchor")
+	if n.Adopted {
+		t.Fatal("converge --yes left the node adopted")
+	}
+	taken, _ := open.inventory.Taken(ctx, "anchor")
+	if !reflect.DeepEqual(taken, []string{"hello-web", overlay.Name, "nftables", "notes"}) {
+		t.Fatalf("the flip took %v", taken)
+	}
+	if !reflect.DeepEqual(*sent, []string{"anchor"}) {
+		t.Fatalf("the flip sent %v", *sent)
+	}
+	declared := composed(t, open, "anchor")
+	if declared.Adoption != nil {
+		t.Fatal("a converged node is still sent an adoption envelope")
+	}
+	if !hasID(declared.Resources, "nftables.filtering") {
+		t.Fatal("the converged node is not declared the mesh's filter")
+	}
+
+	if _, err := adopt(ctx, open, "anchor"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := adopt(ctx, open, "anchor"); err == nil {
+		t.Fatal("adopting an adopted node was not refused")
+	}
+	again, _ := open.inventory.Taken(ctx, "anchor")
+	if !reflect.DeepEqual(again, taken) {
+		t.Fatalf("returning to adopted lost what was taken: %v", again)
+	}
+	declared = composed(t, open, "anchor")
+	if declared.Adoption == nil || len(declared.Adoption.Untaken) != 0 {
+		t.Fatalf("returned to adopted, the envelope is %+v", declared.Adoption)
+	}
+	if hasID(declared.Resources, "nftables.filtering") || hasID(declared.Resources, "nftables.load") {
+		t.Fatal("returned to adopted, the mesh's filter is still declared")
+	}
+}
+
+func hasID(resources []map[string]any, id string) bool {
+	for _, r := range resources {
+		if r["id"] == id {
+			return true
+		}
+	}
+	return false
+}
+
+// The command API refuses a flip exactly as the command line does, in the same words.
+func TestTheApiRefusesTheFlipInTheCommandLinesWords(t *testing.T) {
+	open, _ := anAdoptedAnchor(t)
+	reportsHolding(t, open, heldContainer)
+	_, direct := converge(t.Context(), open, "anchor", true, "", "")
+	if direct == nil {
+		t.Fatal("the flip was not refused")
+	}
+	got := asking(t, letIn{}, "POST", "/converge", `{"node":"anchor","yes":true}`)
+	if got.Code != http.StatusConflict {
+		t.Fatalf("got %d: %s", got.Code, got.Body.String())
+	}
+	var said map[string]any
+	if err := json.Unmarshal(got.Body.Bytes(), &said); err != nil {
+		t.Fatal(err)
+	}
+	if said["refused"] != direct.Error() {
+		t.Fatalf("the API said %q and the command line %q", said["refused"], direct.Error())
+	}
+	if got := asking(t, letIn{}, "POST", "/take", `{"node":"anchor"}`); got.Code != http.StatusBadRequest {
+		t.Fatalf("a take naming no module got %d", got.Code)
+	}
+}
+
+// novox/hq ADR 0100: the preview says what the derived filter does, rendered as it is rendered. On
+// a machine that faces inward, ssh is admitted from the private network only, and a port a module
+// admits from the mesh only closes to everything outside it: both are said to close.
+func TestThePreviewSaysWhatNarrowsToTheMeshCloses(t *testing.T) {
+	open, _ := anAdoptedAnchor(t)
+	ctx := t.Context()
+	register(t, open, catalogue.Manifest{Module: "store", Version: "1",
+		Listens: []catalogue.Listening{{Port: 5432, From: catalogue.FromMesh}}})
+	if _, err := assign(ctx, open, "anchor", "store"); err != nil {
+		t.Fatal(err)
+	}
+	reportsReaching(t, open, []link.Reach{
+		{Protocol: "tcp", Address: "0.0.0.0", Port: 22, By: "sshd"},
+		{Protocol: "tcp", Address: "0.0.0.0", Port: 5432, By: "postgres"},
+		{Protocol: "tcp", Address: "10.77.0.1", Port: 5432, By: "postgres"},
+		{Protocol: "tcp", Address: "0.0.0.0", Port: 8080, By: "hello-web", Published: true,
+			ContainerPort: 80},
+	})
+	preview, err := converge(ctx, open, "anchor", false, "", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	lines := map[string]string{}
+	for _, line := range strings.Split(preview, "\n") {
+		fields := strings.Fields(line)
+		if len(fields) > 1 && strings.HasPrefix(fields[0], "tcp/") {
+			lines[fields[0]+" "+fields[1]] += line + "\n"
+		}
+	}
+	if got := lines["tcp/22 sshd"]; !strings.Contains(got, "WILL CLOSE to everything outside "+
+		"the private network") {
+		t.Errorf("ssh on an inward machine is not said to close outside the mesh:\n%s", preview)
+	}
+	store := lines["tcp/5432 postgres"]
+	if strings.Count(store, "WILL CLOSE to everything outside the private network") != 1 ||
+		!strings.Contains(store, "declared by store (from mesh)") {
+		t.Errorf("the store's narrowing is not said to close, or its mesh address is:\n%s", preview)
+	}
+	if got := lines["tcp/8080 hello-web"]; !strings.Contains(got, "declared by hello-web (from anywhere)") {
+		t.Errorf("a port open to everywhere is not said to stay:\n%s", preview)
+	}
+}
+
+// digestIn is the digest a converge preview printed.
+func digestIn(t *testing.T, preview string) string {
+	t.Helper()
+	for _, line := range strings.Split(preview, "\n") {
+		if fields := strings.Fields(line); len(fields) == 2 && fields[0] == "preview" {
+			return fields[1]
+		}
+	}
+	t.Fatalf("the preview printed no digest:\n%s", preview)
+	return ""
+}
+
+// The flip acts on the preview the operator saw: it names that preview's digest, and it is refused
+// when the digest is missing, when anything the preview says has changed since, or when the node's
+// account of itself is too old to be the machine as it is.
+func TestTheFlipActsOnlyOnThePreviewTheOperatorSaw(t *testing.T) {
+	open, sent := anAdoptedAnchor(t)
+	ctx := t.Context()
+	if _, err := take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true}); err != nil {
+		t.Fatal(err)
+	}
+	reportsHolding(t, open, heldFile)
+	preview, err := converge(ctx, open, "anchor", false, "", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	saw := digestIn(t, preview)
+	if !strings.Contains(preview, "converge anchor --yes "+saw) {
+		t.Fatalf("the preview does not say how to act on it:\n%s", preview)
+	}
+	unchanged := func() {
+		t.Helper()
+		if n, _ := open.inventory.NodeByName(ctx, "anchor"); !n.Adopted || len(*sent) != 0 {
+			t.Fatal("a refused flip changed something")
+		}
+	}
+
+	if _, err := converge(ctx, open, "anchor", true, "", ""); err == nil ||
+		!strings.Contains(err.Error(), "--yes "+saw) {
+		t.Fatalf("a flip naming no preview was not refused: %v", err)
+	}
+	unchanged()
+
+	// Something new is reachable: the preview the operator saw is not what would happen.
+	reportsReaching(t, open, []link.Reach{
+		{Protocol: "tcp", Address: "0.0.0.0", Port: 22, By: "sshd"},
+		{Protocol: "tcp", Address: "0.0.0.0", Port: 8080, By: "hello-web", Published: true,
+			ContainerPort: 80},
+		{Protocol: "tcp", Address: "0.0.0.0", Port: 6000, By: "something-new"},
+	}, heldFile)
+	_, err = converge(ctx, open, "anchor", true, saw, "")
+	if err == nil || !strings.Contains(err.Error(), "has changed since preview "+saw) {
+		t.Fatalf("a flip on a changed preview was not refused: %v", err)
+	}
+	unchanged()
+
+	// An account older than the flip trusts is refused, whatever digest is named.
+	again, err := converge(ctx, open, "anchor", false, "", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	saved := reportFreshFor
+	reportFreshFor = time.Nanosecond
+	_, err = converge(ctx, open, "anchor", true, digestIn(t, again), "")
+	reportFreshFor = saved
+	if err == nil || !strings.Contains(err.Error(), "wait for its next report") {
+		t.Fatalf("a flip on an old account was not refused: %v", err)
+	}
+	unchanged()
+
+	if _, err := converge(ctx, open, "anchor", true, digestIn(t, again), ""); err != nil {
+		t.Fatalf("the flip on the preview just seen was refused: %v", err)
+	}
+	if n, _ := open.inventory.NodeByName(ctx, "anchor"); n.Adopted {
+		t.Fatal("the flip did not converge the node")
+	}
+}
+
+// The flip holds the node from its checks to its send, so a push composed meanwhile waits and is
+// composed after it — never sent after it with the node still adopted. And the send inside the
+// flip is not made to wait on the flip's own hold.
+func TestTheFlipHoldsTheNodeWhileItSends(t *testing.T) {
+	open, _ := anAdoptedAnchor(t)
+	ctx := t.Context()
+	if _, err := take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true}); err != nil {
+		t.Fatal(err)
+	}
+	reportsHolding(t, open, heldFile)
+	preview, err := converge(ctx, open, "anchor", false, "", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	var heldElsewhere, heldHere error
+	sendNodes = func(inner context.Context, open *stores, names []string) error {
+		// Another caller cannot hold the node while the flip sends it.
+		waiting, cancel := context.WithTimeout(ctx, 300*time.Millisecond)
+		defer cancel()
+		if release, err := open.inventory.HoldNodes(waiting, names); err == nil {
+			release()
+			heldElsewhere = errors.New("another caller held the node while the flip sent it")
+		}
+		// The flip's own send holds it without waiting on itself.
+		_, release, err := holdNodes(inner, open, names)
+		if err != nil {
+			heldHere = err
+			return err
+		}
+		release()
+		return nil
+	}
+	if _, err := converge(ctx, open, "anchor", true, digestIn(t, preview), ""); err != nil {
+		t.Fatal(err)
+	}
+	if heldElsewhere != nil || heldHere != nil {
+		t.Fatalf("%v %v", heldElsewhere, heldHere)
+	}
+	// And given back once it is done.
+	release, err := open.inventory.HoldNodes(ctx, []string{"anchor"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	release()
+}
+
+// novox/hq ADR 0103: the preview names every kind of thing a module the flip takes holds as found,
+// not only its files, and the digest changes when any of them does.
+func TestThePreviewNamesEveryHeldKind(t *testing.T) {
+	open, _ := anAdoptedAnchor(t)
+	ctx := t.Context()
+	if _, err := take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true}); err != nil {
+		t.Fatal(err)
+	}
+	since := time.Now()
+	held := []link.Held{heldFile,
+		{ID: "notes.data", Module: "notes", Kind: "directory", Target: "/var/lib/notes", Since: since},
+		{ID: "notes.daemon", Module: "notes", Kind: "service", Target: "notes.service", Since: since},
+		{ID: "notes.seed", Module: "notes", Kind: "archive", Target: "/srv/notes", Since: since},
+		{ID: "notes.worker", Module: "notes", Kind: "process", Target: "notes-worker", Since: since},
+		{ID: "notes.account", Module: "notes", Kind: "user", Target: "notes", Since: since},
+	}
+	reportsHolding(t, open, held...)
+	preview, err := converge(ctx, open, "anchor", false, "", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, want := range []string{
+		"replacing the found directory /var/lib/notes (notes.data)",
+		"replacing the found service notes.service (notes.daemon)",
+		"replacing the found archive /srv/notes (notes.seed)",
+		"replacing the found process notes-worker (notes.worker)",
+		"replacing the found user notes (notes.account)",
+	} {
+		if !strings.Contains(preview, want) {
+			t.Errorf("the preview does not say %q:\n%s", want, preview)
+		}
+	}
+	reportsHolding(t, open, held[:len(held)-1]...)
+	fewer, err := converge(ctx, open, "anchor", false, "", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if digestIn(t, fewer) == digestIn(t, preview) {
+		t.Fatal("the digest does not change with what is held")
+	}
+}
+
+// novox/hq ADR 0100: the flip acts on what the node said is reachable, so an account naming nothing
+// is refused. Every machine that is up answers on ssh; nothing reported means the host's collectors
+// did not, and flipping would close ports the preview never named.
+func TestTheFlipIsRefusedOnAnAccountNamingNothingReachable(t *testing.T) {
+	open, sent := anAdoptedAnchor(t)
+	ctx := t.Context()
+	if _, err := take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true}); err != nil {
+		t.Fatal(err)
+	}
+	// Only a loopback listener: nothing off the machine, which is the same silence.
+	reportsReaching(t, open, []link.Reach{
+		{Protocol: "tcp", Address: "127.0.0.1", Port: 15672, By: "mesh-broker"},
+	}, heldFile)
+	preview, err := converge(ctx, open, "anchor", false, "", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(preview, "this account looks partial") {
+		t.Errorf("the preview does not mark a partial account:\n%s", preview)
+	}
+	_, err = converge(ctx, open, "anchor", true, digestIn(t, preview), "")
+	if err == nil || !strings.Contains(err.Error(), "says nothing is reachable on it") {
+		t.Fatalf("the flip was not refused on an account naming nothing: %v", err)
+	}
+	if n, _ := open.inventory.NodeByName(ctx, "anchor"); !n.Adopted || len(*sent) != 0 {
+		t.Fatal("a refused flip changed something")
+	}
+}
+
+// An assignment cannot land between a preview and the flip that takes every module: assigning
+// holds the node, so it waits for whatever is converging it.
+func TestAssigningWaitsForWhateverIsConvergingTheNode(t *testing.T) {
+	open, _ := anAdoptedAnchor(t)
+	ctx := t.Context()
+	if _, err := take(ctx, open, "anchor", "hello-web", takeOptions{Yes: true}); err != nil {
+		t.Fatal(err)
+	}
+	reportsHolding(t, open, heldFile)
+	preview, err := converge(ctx, open, "anchor", false, "", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	saved, savedPoll := inventory.HoldWaitFor, inventory.HoldPoll
+	inventory.HoldWaitFor, inventory.HoldPoll = time.Second, 50*time.Millisecond
+	defer func() { inventory.HoldWaitFor, inventory.HoldPoll = saved, savedPoll }()
+
+	var whileFlipping error
+	sendNodes = func(context.Context, *stores, []string) error {
+		_, whileFlipping = assign(ctx, open, "anchor", "notes")
+		return nil
+	}
+	if _, err := converge(ctx, open, "anchor", true, digestIn(t, preview), ""); err != nil {
+		t.Fatal(err)
+	}
+	if !errors.Is(whileFlipping, inventory.ErrNodeBusy) {
+		t.Fatalf("an assignment landed while the node was being converged: %v", whileFlipping)
+	}
+}

cmd/mesh-controller/api.go

@@ -0,0 +1,186 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+)
+
+// The command API: what the mesh can be asked to do, over a network.
+//
+// **An adapter and nothing else** (novox/hq ADR 0035). Every route here calls the same function
+// the command line calls, so a refusal is the same refusal in the same words. Nothing is decided
+// in this file — the moment it validates something the command line does not, the mesh has two
+// answers to one question.
+//
+// **It refuses everything unless it was told how to know who is asking.** The board is published
+// on a public name, and this is what stands behind it: an unauthenticated command surface reachable
+// from the internet is authority over the mesh handed to whoever finds it. So there is no
+// permissive default and no flag that removes the check — a mesh that has not been told how to
+// authenticate serves nothing, loudly.
+//
+// The intended authenticator is an OAuth2 provider (ADR 0035), which is an ordinary module. Until
+// one is configured this refuses, which is the correct behaviour rather than a placeholder: a
+// surface that worked without authentication would be one somebody left running.
+func apiCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("api", flag.ContinueOnError)
+	listen := set.String("listen", "127.0.0.1:8081", "where to serve it")
+	issuer := set.String("issuer", "",
+		"the OAuth2 issuer whose tokens this accepts; without it, nothing is served")
+	if _, err := parseAround(set, args); err != nil {
+		return err
+	}
+	if strings.TrimSpace(*issuer) == "" {
+		// Refused at start rather than per request, so it is discovered by whoever ran it rather
+		// than by whoever finds it.
+		return errors.New(
+			"--issuer is not set, and this serves commands rather than pages: it will not run " +
+				"without being told whose tokens to believe. An OAuth2 provider is an ordinary " +
+				"module (novox/hq ADR 0035)")
+	}
+
+	server := &http.Server{
+		Addr:              *listen,
+		ReadHeaderTimeout: 10 * time.Second,
+		Handler:           commands(mustAuthenticate(*issuer)),
+	}
+	fmt.Printf("the command API is on http://%s\n", *listen)
+	fmt.Printf("  it accepts tokens from %s and refuses everything else\n", *issuer)
+	fmt.Printf("  every route calls what the command line calls\n")
+
+	go func() {
+		<-ctx.Done()
+		closing, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+		defer cancel()
+		_ = server.Shutdown(closing)
+	}()
+	if err := server.ListenAndServe(); err != nil && !errors.Is(err, http.ErrServerClosed) {
+		return err
+	}
+	return nil
+}
+
+// Authenticator says whether a request may act, and as whom.
+//
+// An interface so the check can be driven by a test without an identity provider, and so the one
+// real implementation is the only thing that has to be right.
+type Authenticator interface {
+	// Who returns the subject a request is acting as, or an error naming why it may not.
+	Who(r *http.Request) (string, error)
+}
+
+// mustAuthenticate is the real one: a bearer token from the configured issuer.
+//
+// **Not yet verifying the signature**, and it says so rather than pretending. Verification needs
+// the issuer's keys, which needs an identity provider to exist — so this refuses every request
+// until that is built, which is the same answer as having no API at all and is honest about why.
+func mustAuthenticate(issuer string) Authenticator { return notYet{issuer: issuer} }
+
+type notYet struct{ issuer string }
+
+func (n notYet) Who(*http.Request) (string, error) {
+	return "", fmt.Errorf(
+		"this mesh has no way to verify a token from %s yet: the identity provider is a module "+
+			"and none is running. Use the command line, which authenticates through nothing "+
+			"because it is already behind the machine's own login", n.issuer)
+}
+
+// commands is the routing, separate so a test can drive it without a listener.
+func commands(who Authenticator) http.Handler {
+	mux := http.NewServeMux()
+
+	mux.HandleFunc("POST /assign", acting(who, true, func(ctx context.Context, open *stores, in request) (string, error) {
+		return assign(ctx, open, in.Node, splitModules(in.Module)...)
+	}))
+	mux.HandleFunc("POST /unassign", acting(who, true, func(ctx context.Context, open *stores, in request) (string, error) {
+		return unassign(ctx, open, in.Node, splitModules(in.Module)...)
+	}))
+	// Adoption (novox/hq ADR 0100): the same acts as `take`, `converge` and `adopt`.
+	mux.HandleFunc("POST /take", acting(who, true, func(ctx context.Context, open *stores, in request) (string, error) {
+		return take(ctx, open, in.Node, in.Module, takeOptions{Yes: in.Yes, Digest: in.Digest})
+	}))
+	mux.HandleFunc("POST /converge", acting(who, false, func(ctx context.Context, open *stores, in request) (string, error) {
+		return converge(ctx, open, in.Node, in.Yes, in.Digest, in.Filter)
+	}))
+	mux.HandleFunc("POST /adopt", acting(who, false, func(ctx context.Context, open *stores, in request) (string, error) {
+		return adopt(ctx, open, in.Node)
+	}))
+
+	// Anything else is said plainly, because a command surface answering 404 to a verb somebody
+	// expected is indistinguishable from one that is down.
+	mux.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
+		refuse(w, http.StatusNotFound, fmt.Errorf(
+			"%s %s is not something this mesh can be asked; it accepts POST /assign, "+
+				"POST /unassign, POST /take, POST /converge and POST /adopt", r.Method, r.URL.Path))
+	})
+	return mux
+}
+
+type request struct {
+	Node   string `json:"node"`
+	Module string `json:"module"`
+	// Yes, Digest and Filter are converge's and take's: do it rather than preview it, the digest
+	// of the preview it acts on, and (converge) which module loads the mesh's filter.
+	Yes    bool   `json:"yes,omitempty"`
+	Digest string `json:"digest,omitempty"`
+	Filter string `json:"filter,omitempty"`
+}
+
+// acting is the shape every route shares: authenticate, read, act, answer.
+func acting(
+	who Authenticator,
+	needsModule bool,
+	do func(context.Context, *stores, request) (string, error),
+) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		if _, err := who.Who(r); err != nil {
+			refuse(w, http.StatusUnauthorized, err)
+			return
+		}
+		var in request
+		if err := json.NewDecoder(r.Body).Decode(&in); err != nil {
+			refuse(w, http.StatusBadRequest, fmt.Errorf("this is not a request this understands: %w", err))
+			return
+		}
+		if in.Node == "" || (needsModule && in.Module == "") {
+			if needsModule {
+				refuse(w, http.StatusBadRequest, errors.New(`both "node" and "module" are needed`))
+			} else {
+				refuse(w, http.StatusBadRequest, errors.New(`"node" is needed`))
+			}
+			return
+		}
+
+		open, err := openStores(r.Context())
+		if err != nil {
+			refuse(w, http.StatusServiceUnavailable, err)
+			return
+		}
+		defer open.Close()
+
+		said, err := do(r.Context(), open, in)
+		if err != nil {
+			// **The refusal the command line would have given, unchanged.** Carrying `said` with
+			// it matters: an assignment that was kept and still does not resolve is two facts,
+			// and dropping either makes the answer wrong.
+			answer(w, http.StatusConflict, map[string]any{"said": said, "refused": err.Error()})
+			return
+		}
+		answer(w, http.StatusOK, map[string]any{"said": said})
+	}
+}
+
+func answer(w http.ResponseWriter, status int, body map[string]any) {
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(status)
+	_ = json.NewEncoder(w).Encode(body)
+}
+
+func refuse(w http.ResponseWriter, status int, err error) {
+	answer(w, status, map[string]any{"refused": err.Error()})
+}

cmd/mesh-controller/api_test.go

@@ -0,0 +1,77 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+type letIn struct{}
+
+func (letIn) Who(*http.Request) (string, error) { return "somebody", nil }
+
+func asking(t *testing.T, who Authenticator, method, path, body string) *httptest.ResponseRecorder {
+	t.Helper()
+	recorded := httptest.NewRecorder()
+	commands(who).ServeHTTP(recorded, httptest.NewRequest(method, path, strings.NewReader(body)))
+	return recorded
+}
+
+// **Nothing is served to somebody the mesh cannot identify**, and that is the default rather than
+// a setting. The board this stands behind is published on a public name (novox/hq 11-a-board), so
+// an unauthenticated command surface is authority over the mesh handed to whoever finds it.
+func TestAnUnauthenticatedRequestIsRefused(t *testing.T) {
+	got := asking(t, mustAuthenticate("https://identity.example/realms/mesh"),
+		"POST", "/assign", `{"node":"anchor","module":"umami"}`)
+	if got.Code != http.StatusUnauthorized {
+		t.Fatalf("an unidentified caller got %d", got.Code)
+	}
+	// And it says what to do instead, because the answer is not "give up".
+	if !strings.Contains(got.Body.String(), "command line") {
+		t.Errorf("the refusal does not say what still works: %s", got.Body.String())
+	}
+}
+
+// The API refuses to start at all without being told whose tokens to believe.
+//
+// At start rather than per request, so it is found by whoever ran it rather than by whoever
+// finds it.
+func TestTheApiWillNotRunWithoutAnIssuer(t *testing.T) {
+	err := apiCommand(context.Background(), []string{})
+	if err == nil {
+		t.Fatal("it served commands without being told who may give them")
+	}
+	if !strings.Contains(err.Error(), "issuer") {
+		t.Errorf("the refusal does not name what is missing: %v", err)
+	}
+}
+
+// A request missing what it acts on is refused before anything is opened.
+func TestARequestThatNamesNothingIsRefused(t *testing.T) {
+	// **The exact refusal, not merely "not accepted".** Asserting non-200 passes even when the
+	// request got as far as opening a store and failing there, which proves nothing about whether
+	// anything was checked — that is what the first version of this test did.
+	for _, body := range []string{`{}`, `{"node":"anchor"}`, `{"module":"umami"}`, `not json`} {
+		got := asking(t, letIn{}, "POST", "/assign", body)
+		if got.Code != http.StatusBadRequest {
+			t.Errorf("%s got %d, and a request naming nothing is a bad request", body, got.Code)
+		}
+	}
+}
+
+// A verb nobody implemented is said plainly. A command surface answering 404 to something
+// somebody expected is indistinguishable from one that is down.
+func TestAnUnknownRouteSaysWhatIsAccepted(t *testing.T) {
+	got := asking(t, letIn{}, "POST", "/rotate", `{}`)
+	if got.Code != http.StatusNotFound {
+		t.Fatalf("got %d", got.Code)
+	}
+	var said map[string]any
+	_ = json.Unmarshal(got.Body.Bytes(), &said)
+	if !strings.Contains(said["refused"].(string), "/assign") {
+		t.Errorf("it does not say what it does accept: %v", said)
+	}
+}

cmd/mesh-controller/ask.go

@@ -0,0 +1,62 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"os"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// ask calls one of a module's tools, through the control plane's own broker connection.
+//
+// A module serves tools under an account scoped to exactly that (novox/hq ADR 0047), and nothing
+// else in the mesh held an account that could ask one — not an operator at a terminal, not an agent
+// acting for one (novox/hq 04-ISSUES/049). The control plane does, so it is the way in: one
+// process, one connection, one place a question can be seen to have been asked (ADR 0095).
+func askCommand(ctx context.Context, args []string) error {
+	positionals, flags := split(args)
+	set := flag.NewFlagSet("ask", flag.ContinueOnError)
+	wait := set.Duration("wait", 60*time.Second, "how long to wait for the module's answer")
+	if err := set.Parse(flags); err != nil {
+		return err
+	}
+	if len(positionals) < 2 || len(positionals) > 3 {
+		return errors.New("ask <module> <tool> [json arguments] [--wait 60s]")
+	}
+	module, tool := positionals[0], positionals[1]
+	var arguments json.RawMessage
+	if len(positionals) == 3 {
+		if !json.Valid([]byte(positionals[2])) {
+			return fmt.Errorf("the arguments are not JSON: %s", positionals[2])
+		}
+		arguments = json.RawMessage(positionals[2])
+	}
+
+	server, err := connectLink(ctx, nil, nil, nil)
+	if err != nil {
+		return err
+	}
+	defer server.Close()
+
+	answer, err := link.Ask(ctx, server.Bus(), module, tool, arguments, *wait)
+	if err != nil {
+		return err
+	}
+	// The answer as the module gave it, to standard output, for a person or a program. A tool
+	// that answered with an error has still answered: printed the same way, and the exit status
+	// says which.
+	body, err := json.Marshal(answer)
+	if err != nil {
+		return err
+	}
+	fmt.Fprintln(os.Stdout, string(body))
+	if answer.Error != "" {
+		return fmt.Errorf("%s.%s answered with an error: %s", module, tool, answer.Error)
+	}
+	return nil
+}

cmd/mesh-controller/board.go

@@ -0,0 +1,324 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"html/template"
+	"net/http"
+	"sort"
+	"strings"
+	"time"
+)
+
+// boardCommand serves the three questions as a page.
+//
+// **It reads through the same functions everything else does and holds nothing**
+// (novox/hq 03-DESIGN/01-to-be/11-a-board.md). The board being replaced is one service that reads
+// every context's database directly — [ADR 0008](novox/hq) violated by the one component with a
+// reason to violate it, and the cost is that a boundary nothing may cross can move, while one
+// thing crossing it is enough to freeze it. A board that reads the provisioning tables is a board
+// that breaks when provisioning changes its tables, and the change then gets weighed against the
+// board.
+//
+// **It stores nothing of its own.** No cache that can disagree, no table of what the mesh looked
+// like last time. Every request reads the mesh now; if that is slow, the answer belongs in the
+// context that owns it, where everything else asking gets it too.
+//
+// **Reading is the whole of it.** Every action a board could offer already exists as a command,
+// and a button that does something no command does is a second implementation of a decision.
+func boardCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("board", flag.ContinueOnError)
+	// The private network, not everything. A board says which machines are broken and what they
+	// are running, which is exactly the map somebody attacking this would like — and there is no
+	// reason for it to be reachable from further away than the mesh.
+	listen := set.String("listen", "127.0.0.1:8080", "where to serve it")
+	if _, err := parseAround(set, args); err != nil {
+		return err
+	}
+
+	server := &http.Server{
+		Addr:              *listen,
+		ReadHeaderTimeout: 10 * time.Second,
+		Handler:           board(),
+	}
+	fmt.Printf("the board is on http://%s\n", *listen)
+	fmt.Printf("  it reads the mesh on every request and keeps nothing\n")
+
+	go func() {
+		<-ctx.Done()
+		closing, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+		defer cancel()
+		_ = server.Shutdown(closing)
+	}()
+	if err := server.ListenAndServe(); err != nil && !errors.Is(err, http.ErrServerClosed) {
+		return err
+	}
+	return nil
+}
+
+// board is the handler, separate so a test can drive it without a listener.
+func board() http.Handler {
+	mux := http.NewServeMux()
+	mux.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/" {
+			http.NotFound(w, r)
+			return
+		}
+		asked, err := ask(r.Context())
+		if err != nil {
+			// **Said, not blank.** A board that cannot reach the mesh and renders an empty page
+			// says "nothing is wrong" in the one situation where nobody can know that.
+			w.Header().Set("Content-Type", "text/html; charset=utf-8")
+			w.WriteHeader(http.StatusServiceUnavailable)
+			_ = page.Execute(w, view{Unreachable: err.Error()})
+			return
+		}
+		w.Header().Set("Content-Type", "text/html; charset=utf-8")
+		if err := page.Execute(w, asked); err != nil {
+			// The page is half-written by now; there is nothing useful left to say to the
+			// browser, and saying it here is what stops the failure being silent.
+			fmt.Printf("the board could not render: %v\n", err)
+		}
+	})
+	// The same answers for something that is not a person, from the same read. A board and a
+	// script disagreeing about which machine is broken would be worse than either alone.
+	mux.HandleFunc("/mesh.json", func(w http.ResponseWriter, r *http.Request) {
+		open, err := openStores(r.Context())
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusServiceUnavailable)
+			return
+		}
+		defer open.Close()
+		asked, err := theThreeQuestions(r.Context(), open)
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusServiceUnavailable)
+			return
+		}
+		body, err := statusAsJSON(asked)
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusInternalServerError)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write(append(body, '\n'))
+	})
+	return mux
+}
+
+// ask reads the mesh for one request.
+func ask(ctx context.Context) (view, error) {
+	open, err := openStores(ctx)
+	if err != nil {
+		return view{}, err
+	}
+	defer open.Close()
+
+	asked, err := theThreeQuestions(ctx, open)
+	if err != nil {
+		return view{}, err
+	}
+	return viewOf(asked), nil
+}
+
+// view is what the page is given. Nothing is derived here that the reader could not derive.
+type view struct {
+	Unreachable string
+	Machines    int
+	Broken      []brokenMachine
+	Quiet       []quietMachine
+	Behind      []staleModule
+	Waiting     []waitingMachine
+	// Unresolved is every machine that cannot be worked out at all. Shown above everything else,
+	// because a machine here is in none of the other lists: nothing was computed for it, so it is
+	// not broken, not quiet and not behind — and a page without this said "all well" about a mesh
+	// where nothing could be sent anywhere.
+	Unresolved []blockedMachine
+	// Network is why the private network could not be computed, when it could not.
+	Network string
+	// Adopted is every node still adopted (novox/hq ADR 0100). Not broken: nothing forces the
+	// flip, so a node left adopted is shown rather than read as converged.
+	Adopted []string
+	At      string
+}
+
+type blockedMachine struct {
+	Node string
+	// Said is the mesh's own words, a line at a time. Every unmet requirement, not the first:
+	// a machine is usually blocked by more than one and fixing one of them changes nothing.
+	Said []string
+}
+
+type waitingMachine struct {
+	Node string
+	// Never told is not out of date: nobody has ever asked this machine to be anything. Same
+	// remedy, different situation, and the page says which.
+	Never bool
+}
+
+type brokenMachine struct {
+	Node string
+	// Outcome is refused or failed, and stays distinct all the way to the page. **Refused means
+	// the machine is exactly as it was and what is wrong is in what was sent; failed means it is
+	// in a state nobody declared and what is wrong is on the machine.** They are fixed in
+	// different places, so one word for both would send half the readers to the wrong one.
+	Outcome string
+	Said    []string
+	When    string
+}
+
+type quietMachine struct {
+	Node string
+	// Heard is "never" or how long ago. Never heard from is not the same as quiet for a while:
+	// one may be a machine that was never sent anything.
+	Heard string
+}
+
+type staleModule struct {
+	Module  string
+	Holds   string
+	Source  string
+	Running []string
+}
+
+func viewOf(asked answers) view {
+	out := view{Machines: len(asked.nodes), At: time.Now().Format("15:04:05"),
+		Network: asked.network, Adopted: adoptedNodes(asked.nodes)}
+	var blocked []string
+	for name := range asked.refused {
+		blocked = append(blocked, name)
+	}
+	sort.Strings(blocked)
+	for _, name := range blocked {
+		one := blockedMachine{Node: name}
+		for _, line := range strings.Split(strings.TrimRight(asked.refused[name], "\n"), "\n") {
+			one.Said = append(one.Said, strings.TrimSpace(line))
+		}
+		out.Unresolved = append(out.Unresolved, one)
+	}
+	for _, d := range asked.wrong {
+		one := brokenMachine{Node: d.Node, Outcome: d.Outcome,
+			When: d.At.Local().Format("2006-01-02 15:04")}
+		if d.Refused != "" {
+			// The host's own words. It says exactly what it could not accept, and nothing
+			// written here would say it better.
+			one.Said = append(one.Said, firstLine(d.Refused))
+		}
+		for _, f := range d.Failed {
+			one.Said = append(one.Said, f.ID+": "+firstLine(f.Error))
+		}
+		out.Broken = append(out.Broken, one)
+	}
+	for _, n := range asked.quiet {
+		out.Quiet = append(out.Quiet, quietMachine{Node: n.Name, Heard: heardFrom(n)})
+	}
+	for _, m := range asked.waiting {
+		out.Waiting = append(out.Waiting, waitingMachine{Node: m.Node, Never: m.Never})
+	}
+	for module, on := range asked.behind {
+		from := asked.sources[module]
+		out.Behind = append(out.Behind, staleModule{
+			Module: module, Holds: short(from.BuiltFrom), Source: short(from.Head), Running: on,
+		})
+	}
+	return out
+}
+
+// The page. Deliberately one file with no assets: a board that cannot render without fetching
+// something is a board that is blank exactly when the mesh is unwell.
+var page = template.Must(template.New("board").Parse(`<!doctype html>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>the mesh</title>
+<style>
+  :root { color-scheme: light dark; }
+  body { font: 15px/1.5 ui-monospace, SFMono-Regular, Menlo, monospace; margin: 2rem auto;
+         max-width: 52rem; padding: 0 1rem; }
+  h1 { font-size: 1.1rem; font-weight: 600; margin: 0 0 1.5rem; }
+  h2 { font-size: 1rem; font-weight: 600; margin: 2rem 0 .5rem; }
+  .quiet { opacity: .65; }
+  .said { opacity: .8; padding-left: 1.5rem; }
+  .outcome { display: inline-block; min-width: 4.5rem; }
+  .refused { color: #b26b00; }
+  .failed { color: #c0392b; }
+  ul { list-style: none; padding: 0; margin: 0; }
+  li { padding: .15rem 0; }
+  footer { margin-top: 3rem; opacity: .6; font-size: .85rem; }
+</style>
+{{if .Unreachable}}
+<h1>the mesh cannot be read</h1>
+<p class="failed">{{.Unreachable}}</p>
+<p class="quiet">This says nothing about whether the mesh is well — only that this page could not
+find out.</p>
+{{else}}
+<h1>{{.Machines}} machine{{if ne .Machines 1}}s{{end}}</h1>
+
+{{if .Unresolved}}
+<h2>Can everything be worked out?</h2>
+<ul>
+  {{range .Unresolved}}
+  <li><span class="outcome failed">blocked</span> <strong>{{.Node}}</strong>
+    {{range .Said}}<div class="said">{{.}}</div>{{end}}
+  </li>
+  {{end}}
+</ul>
+<p class="quiet">Nothing can be sent to a machine here, and it appears in none of the lists below:
+nothing was computed for it, so there is nothing it can be behind.</p>
+{{end}}
+{{if .Network}}
+<p class="failed">The private network could not be computed: {{.Network}}</p>
+{{end}}
+
+<h2>Is anything broken?</h2>
+{{if .Broken}}
+<ul>
+  {{range .Broken}}
+  <li>
+    <span class="outcome {{.Outcome}}">{{.Outcome}}</span>
+    <strong>{{.Node}}</strong> <span class="quiet">{{.When}}</span>
+    {{range .Said}}<div class="said">{{.}}</div>{{end}}
+  </li>
+  {{end}}
+</ul>
+{{else}}<p class="quiet">No. Every machine is doing what it was told.</p>{{end}}
+
+<h2>Is anything not answering?</h2>
+{{if .Quiet}}
+<ul>{{range .Quiet}}<li><strong>{{.Node}}</strong> <span class="quiet">{{.Heard}}</span></li>{{end}}</ul>
+<p class="quiet">Not heard from is not the same as tried and could not — a machine here may be
+new, switched off, or unreachable.</p>
+{{else}}<p class="quiet">No. Every machine has been heard from.</p>{{end}}
+
+<h2>Is anything out of date?</h2>
+{{if .Behind}}
+<ul>
+  {{range .Behind}}
+  <li><strong>{{.Module}}</strong> holds {{.Holds}}, source has {{.Source}}
+    {{if .Running}}<div class="said">running on {{range $i, $n := .Running}}{{if $i}}, {{end}}{{$n}}{{end}}</div>
+    {{else}}<div class="said quiet">assigned to nothing</div>{{end}}
+  </li>
+  {{end}}
+</ul>
+{{else}}<p class="quiet">No. Every module is what its source last had.</p>{{end}}
+
+{{if .Waiting}}
+<ul>
+  {{range .Waiting}}
+  <li><strong>{{.Node}}</strong>
+    {{if .Never}}<span class="quiet">has never been sent anything</span>
+    {{else}}<span class="quiet">is not running what the mesh would send it</span>{{end}}
+  </li>
+  {{end}}
+</ul>
+<p class="quiet">Never told is not out of date: nobody has asked that machine to be anything yet.
+Both are sent by <code>push --behind</code>.</p>
+{{else}}<p class="quiet">Every machine is running what the mesh would send it.</p>{{end}}
+{{if .Adopted}}
+<h2>Which machines are adopted?</h2>
+<ul>{{range .Adopted}}<li><strong>{{.}}</strong> <span class="quiet">adopted — what was found on it is kept until each module is taken</span></li>{{end}}</ul>
+{{end}}
+{{end}}
+
+<footer>Read at {{.At}}. This page holds nothing and changes nothing.</footer>
+`))

cmd/mesh-controller/board_test.go

@@ -0,0 +1,154 @@
+package main
+
+import (
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// Refused and failed stay distinct all the way to the page.
+//
+// **Refused means the machine is exactly as it was and what is wrong is in what was sent; failed
+// means it is in a state nobody declared and what is wrong is on the machine.** They are fixed in
+// different places, so a page saying "error" for both sends half its readers to the wrong one.
+func TestRefusedAndFailedReachThePageAsDifferentThings(t *testing.T) {
+	rendered := render(t, viewOf(answers{
+		nodes: []inventory.Node{{Name: "anchor"}, {Name: "laptop"}},
+		wrong: []inventory.Doing{
+			{Node: "anchor", Outcome: "refused", Refused: "not a declaration this host speaks",
+				At: time.Now()},
+			{Node: "laptop", Outcome: "failed", At: time.Now(),
+				Failed: []inventory.FailedResource{{ID: "web.container", Error: "no such image"}}},
+		},
+	}))
+	// The rendered outcome, not the word anywhere on the page: both words appear in the
+	// stylesheet, so a plain Contains passes whatever the machine actually said. It did.
+	for _, want := range []string{`<span class="outcome refused">refused</span>`,
+		`<span class="outcome failed">failed</span>`} {
+		if !strings.Contains(rendered, want) {
+			t.Fatalf("the page does not distinguish the two — missing %s:\n%s", want, rendered)
+		}
+	}
+	// And the host's own words, which say exactly what it could not accept.
+	for _, want := range []string{"not a declaration this host speaks", "web.container", "no such image"} {
+		if !strings.Contains(rendered, want) {
+			t.Fatalf("the page does not say %q, so a reader must go and ask:\n%s", want, rendered)
+		}
+	}
+}
+
+// Nothing wrong is said, not left blank. An empty page and a well mesh must not look alike.
+func TestAWellMeshSaysSoRatherThanShowingNothing(t *testing.T) {
+	rendered := render(t, viewOf(answers{nodes: []inventory.Node{{Name: "anchor"}}}))
+	for _, want := range []string{
+		"Every machine is doing what it was told",
+		"Every machine has been heard from",
+		"Every module is what its source last had",
+	} {
+		if !strings.Contains(rendered, want) {
+			t.Fatalf("a well mesh does not say %q:\n%s", want, rendered)
+		}
+	}
+}
+
+// A board that cannot reach the mesh must not render an empty page: that says "nothing is wrong"
+// in the one situation where nobody can know it.
+func TestABoardThatCannotReadTheMeshSaysSo(t *testing.T) {
+	rendered := render(t, view{Unreachable: "the store did not answer"})
+	if !strings.Contains(rendered, "the store did not answer") {
+		t.Fatalf("the reason is not on the page:\n%s", rendered)
+	}
+	if strings.Contains(rendered, "Every machine is doing what it was told") {
+		t.Fatal("a board that could not read the mesh reported that the mesh is well")
+	}
+}
+
+// Quiet is not broken, and the page says which it is.
+func TestQuietIsNotReportedAsBroken(t *testing.T) {
+	rendered := render(t, viewOf(answers{
+		nodes: []inventory.Node{{Name: "laptop"}},
+		quiet: []inventory.Node{{Name: "laptop"}},
+	}))
+	if !strings.Contains(rendered, "not the same as tried and could not") {
+		t.Fatalf("the page does not separate quiet from broken:\n%s", rendered)
+	}
+	if !strings.Contains(rendered, "Every machine is doing what it was told") {
+		t.Fatalf("a quiet machine was counted as broken:\n%s", rendered)
+	}
+}
+
+// The page renders what a machine said, and a hostile string in it is not markup.
+//
+// What is on this page comes from machines, and a machine's own words are the whole reason the
+// page is useful. They are also the one thing here that nobody in this repository wrote.
+func TestWhatAMachineSaidIsNotMarkup(t *testing.T) {
+	rendered := render(t, viewOf(answers{
+		nodes: []inventory.Node{{Name: "anchor"}},
+		wrong: []inventory.Doing{{Node: "anchor", Outcome: "refused",
+			Refused: `<script>alert("from the machine")</script>`, At: time.Now()}},
+	}))
+	if strings.Contains(rendered, "<script>alert") {
+		t.Fatalf("a machine's words were rendered as markup:\n%s", rendered)
+	}
+	if !strings.Contains(rendered, "&lt;script&gt;") {
+		t.Fatalf("the words were not shown at all, so the reader cannot see what it said:\n%s", rendered)
+	}
+}
+
+// A module behind its source names the machines running the old one — the part with consequences.
+func TestBeingBehindNamesTheMachinesRunningTheOldOne(t *testing.T) {
+	rendered := render(t, viewOf(answers{
+		nodes:   []inventory.Node{{Name: "laptop"}},
+		behind:  map[string][]string{"web": {"laptop"}},
+		sources: map[string]inventory.Source{"web": {BuiltFrom: "aaaaaaaaaa", Head: "bbbbbbbbbb"}},
+	}))
+	for _, want := range []string{"web", "aaaaaaaa", "bbbbbbbb", "running on laptop"} {
+		if !strings.Contains(rendered, want) {
+			t.Fatalf("the page does not say %q:\n%s", want, rendered)
+		}
+	}
+}
+
+func render(t *testing.T, v view) string {
+	t.Helper()
+	var out strings.Builder
+	if err := page.Execute(&out, v); err != nil {
+		t.Fatal(err)
+	}
+	return out.String()
+}
+
+// The board and the command must answer the same question the same way.
+//
+// A machine not running what the mesh would send it was added to `status` and would have been
+// missing here — which is the one thing this page's design forbids: two answers to the same
+// question, disagreeing, with a person in front of each.
+func TestTheBoardSaysWhichMachinesHaveNotBeenSentWhatTheyShouldBe(t *testing.T) {
+	rendered := render(t, viewOf(answers{
+		nodes: []inventory.Node{{Name: "anchor"}, {Name: "laptop"}},
+		waiting: []inventory.Machine{
+			{Node: "anchor"},
+			{Node: "laptop", Never: true},
+		},
+	}))
+	if !strings.Contains(rendered, "is not running what the mesh would send it") {
+		t.Fatalf("the page does not say a machine is out of date:\n%s", rendered)
+	}
+	// Never told is a different situation with the same remedy, and the page keeps them apart.
+	if !strings.Contains(rendered, "has never been sent anything") {
+		t.Fatalf("the page does not separate never told from out of date:\n%s", rendered)
+	}
+	if !strings.Contains(rendered, "push --behind") {
+		t.Fatalf("the page does not say what sends them:\n%s", rendered)
+	}
+}
+
+// And says so when there is nothing waiting, rather than leaving the question unanswered.
+func TestAMeshWithNothingWaitingSaysSo(t *testing.T) {
+	rendered := render(t, viewOf(answers{nodes: []inventory.Node{{Name: "anchor"}}}))
+	if !strings.Contains(rendered, "Every machine is running what the mesh would send it") {
+		t.Fatalf("the page leaves the question unanswered:\n%s", rendered)
+	}
+}

cmd/mesh-controller/build.go

@@ -0,0 +1,834 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"os"
+	"strings"
+	"time"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// asking a build machine for a module, and what came back.
+//
+// Split out of main.go, which had reached 2,769 lines because appending was always the
+// cheapest next step. That is how novox/hq ADR 0001 records `hal/sdk` reaching 34,636:
+// nothing in it was wrong, and no one edit was the one that should have been a new file.
+
+// buildCommand builds a module from its source and records what came out.
+//
+// **Run where there is a container runtime**, which is why it is a command rather than something
+// the control plane does on its own: building needs to run things on a machine, and what the
+// control plane may send a machine is bounded by the declaration language. This is the shape the
+// builder module will take when it is given work over the broker; today a person runs it, and the
+// mesh records the result the same way either way.
+// buildOn rebuilds every module the mesh holds that stands on the named module's artifacts — the
+// rebuild a changed base needs, which nothing else asks for: their sources did not move, and
+// "behind" does not see a base that did (novox/hq 04-ISSUES/131). Bases first among them too.
+func buildOn(ctx context.Context, base string, wait time.Duration) error {
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	held, err := open.inventory.Catalogued(ctx)
+	if err != nil {
+		return err
+	}
+	against, err := open.inventory.BuiltAgainst(ctx)
+	if err != nil {
+		return err
+	}
+	var on []inventory.Entry
+	for _, e := range held {
+		if standsOnModule(e, base, against) {
+			on = append(on, e)
+		}
+	}
+	if len(on) == 0 {
+		fmt.Printf("nothing the mesh holds stands on %s\n", base)
+		return nil
+	}
+	on = orderByBases(on, against)
+	fmt.Printf("%d module(s) stand on %s:\n", len(on), base)
+	var failed []string
+	for _, e := range on {
+		fmt.Printf("--- %s\n", e.Manifest.Module)
+		source := buildSource{Repository: e.Source.Repository, Seat: e.Source.Seat}
+		if err := buildOne(ctx, source, e.Source.Path, e.Source.Ref, wait); err != nil {
+			fmt.Printf("  %v\n", err)
+			failed = append(failed, e.Manifest.Module)
+		}
+	}
+	if len(failed) > 0 {
+		return fmt.Errorf("%d of %d could not be built: %s", len(failed), len(on), strings.Join(failed, ", "))
+	}
+	fmt.Printf("\n%d module(s) rebuilt on %s. `push --behind` sends them on\n", len(on), base)
+	return nil
+}
+
+func buildCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("build", flag.ContinueOnError)
+	ref := set.String("ref", "", "the branch, tag or commit to build")
+	// A module is a repository and a path within it (novox/hq ADR 0069). Empty is the repository's
+	// root, which is the ordinary case and why this is a flag rather than a second argument.
+	path := set.String("path", "", "the module's directory inside the repository")
+	wait := set.Duration("wait", 10*time.Minute, "how long to wait for a builder to answer")
+	dryRun := set.Bool("dry-run", false, "build and print the manifest, recording nothing")
+	// Every module whose source has moved, rather than one named repository.
+	//
+	// **The mirror of `push --behind`, and the same argument** (novox/hq ADR 0010): the mesh
+	// already knows which modules are behind their source, so making a person read that list and
+	// retype each repository is asking them to be the loop. Naming a repository and asking which
+	// ones need building are different requests, so they are not combined.
+	behind := set.Bool("behind", false, "every module the mesh holds older than its source has")
+	on := set.String("on", "", "rebuild every module that stands on this module's artifacts — the rebuild a changed base needs")
+	// A repository on the mesh's own forge, named by its path there (novox/hq ADR 0111). Without it
+	// the repository is external, cloned exactly as given — see source.go.
+	self := set.Bool("self", false, "the repository is a path on the forge holding the git seat")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if *on != "" {
+		if len(positionals) != 0 || *behind || *self {
+			return errors.New("build --on <module> names a base and nothing else")
+		}
+		return buildOn(ctx, *on, *wait)
+	}
+
+	if *behind {
+		if len(positionals) != 0 || *self {
+			return errors.New("build <repository> or build --behind, not both: one names a " +
+				"repository and the other asks which need building")
+		}
+		return buildBehind(ctx, *wait)
+	}
+	if len(positionals) != 1 {
+		return errors.New("build <repository> [--self] [--path P] [--ref R] [--wait D] [--dry-run]  |  build --behind  |  build --on <module>")
+	}
+	source := buildSource{Repository: positionals[0]}
+	if *self {
+		if err := onASeat(source.Repository); err != nil {
+			return err
+		}
+		source.Seat = gitSeat
+	}
+
+	if *dryRun {
+		return buildAndShow(ctx, source, *path, *ref, *wait)
+	}
+	return buildOne(ctx, source, *path, *ref, *wait)
+}
+
+// buildFrom turns what a builder said into what the mesh keeps.
+//
+// **By digest and path, never by where it was pushed** (novox/hq 04-ISSUES/102). The builder
+// says `<registry>:<port>/<module>/<artifact>@sha256:…`; the mesh records the artifact-store
+// reference and composes the store's address back in where a reference is used. `against` — what
+// the build stood on, the catalogue's edge — is recorded the same way, so an edge names a module's
+// artifact and not the machine it was pulled from.
+func buildFrom(result link.BuildResult) inventory.Build {
+	kept := inventory.Build{
+		ID: result.ID, Repository: result.Repository, Ref: result.Ref,
+		Commit: result.Commit, On: result.On, Failed: result.Failed,
+		// **What the announcement carries, kept rather than discarded** (novox/hq 04-ISSUES/050).
+		// The catalogue turns the manifest into requires/provides edges and `against` into build
+		// edges, and it is not always listening when a build happens — on a fresh mesh it cannot
+		// be, for exactly the modules it needs most. Keeping them is what makes a replay able to
+		// rebuild the graph rather than a list of names.
+		Path: result.Path,
+	}
+	// When it was asked, which is what orders it against another build of the same module
+	// (novox/hq 04-ISSUES/219) — not when it was heard.
+	if asked, ok := link.BuildAskedAt(result.ID); ok {
+		kept.Asked = asked
+	}
+	for _, ref := range result.Against {
+		kept.Against = append(kept.Against, catalogue.Recorded(ref))
+	}
+	for _, r := range result.Read {
+		kept.Read = append(kept.Read, inventory.ReadRepository{Repository: r.Repository, Ref: r.Ref})
+	}
+	var announced []inventory.Artifact
+	for _, made := range result.Made {
+		announced = append(announced, inventory.Artifact{
+			Name: made.Name, Kind: made.Kind, Reference: made.Reference,
+		})
+		kept.Made = append(kept.Made, inventory.Artifact{
+			Name: made.Name, Kind: made.Kind, Reference: catalogue.Recorded(made.Reference),
+		})
+	}
+	kept.Manifest = recordedManifest(result.Manifest, announced)
+	// The module name comes from the manifest, which only exists when the build got that far.
+	if len(kept.Manifest) > 0 {
+		if m, err := catalogue.ParseManifest(kept.Manifest); err == nil {
+			kept.Module = m.Module
+		}
+	}
+	return kept
+}
+
+// buildsCommand says what has been built lately.
+func buildsCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("builds", flag.ContinueOnError)
+	limit := set.Int("n", 20, "how many to show")
+	logOf := set.String("log", "", "a build's id: print what the build machine said, line by line")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if *logOf != "" {
+		return buildLog(ctx, *logOf)
+	}
+	module := ""
+	if len(positionals) == 1 {
+		module = positionals[0]
+	} else if len(positionals) > 1 {
+		return errors.New("builds [<module>] [-n N]")
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	builds, err := inv.Builds(ctx, module, *limit)
+	if err != nil {
+		return err
+	}
+	if len(builds) == 0 {
+		// Said rather than printed as nothing: an empty list and a failed read must never look
+		// the same, and getting here means the store answered.
+		if module != "" {
+			fmt.Printf("nothing has been built for %s\n", module)
+			return nil
+		}
+		fmt.Println("nothing has been built yet")
+		return nil
+	}
+
+	for _, b := range builds {
+		what := b.Module
+		if what == "" {
+			// It failed before knowing what it was building, which is most of the interesting
+			// failures. The repository is what a person has to go and look at.
+			what = "?"
+		}
+		outcome := "built " + short(b.Commit)
+		if !b.Worked() {
+			outcome = "failed"
+		}
+		fmt.Printf("%-18s %-14s %-10s %s  %s\n",
+			what, outcome, b.On, b.At.Local().Format("2006-01-02 15:04"), b.ID)
+		fmt.Printf("    %s", b.Repository)
+		if b.Ref != "" {
+			fmt.Printf(" at %s", b.Ref)
+		}
+		fmt.Println()
+		for _, made := range b.Made {
+			fmt.Printf("    %-10s %s\n", made.Kind, made.Reference)
+		}
+		if !b.Worked() {
+			// The builder's own first line. The whole failure is often a build log, and printing
+			// it here would bury every other row.
+			fmt.Printf("    %s\n", firstLine(b.Failed))
+		}
+	}
+	return nil
+}
+
+// builderCommand issues a build machine its own broker credential.
+//
+// **A build machine is not a node**, and giving it a node's account would let it read another
+// machine's declarations. This is narrower and different: read the build queue, write the
+// exchange and an asker's reply queue, and nothing else.
+//
+// Issued rather than assumed, because until this the builder used whatever credential it was
+// handed — which in practice meant the broker's own administrative one. A program documented as
+// holding its own credential and given somebody else's is worse than one with no story at all.
+func builderCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("builder issue", flag.ContinueOnError)
+	// Which machine will use it. Given, the credential is delivered by the mesh rather than
+	// printed for somebody to carry — which is the difference between the builder being a module
+	// and being a program somebody configures.
+	forNode := set.String("node", "",
+		"the machine that will run it, so the mesh delivers the credential instead of printing it")
+	module := set.String("module", "builder", "the module on that machine that will read it")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 2 || positionals[0] != "issue" {
+		return errors.New("builder issue <name> [--node <machine>]")
+	}
+	name := positionals[1]
+
+	// **The build machine's credential is a module's credential** (novox/hq ADR 0131, design 28
+	// task 5.5): minted into the mesh's records and sealed to the machine as the builder module's
+	// broker secret, usable at the next push — the same act `module issue` performs, and the same
+	// account the composed user list carries. Nothing is created on a server; the bus reads the list.
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return err
+	}
+	m, known := shelf[*module]
+	if !known {
+		return fmt.Errorf("%s is not in the catalogue; `module add` it first", *module)
+	}
+	fmt.Printf("build machine %s: ", name)
+	node := *forNode
+	if node == "" {
+		entries, err := inv.Catalogued(ctx)
+		if err != nil {
+			return err
+		}
+		for _, e := range entries {
+			if e.Manifest.Module == *module && len(e.On) > 0 {
+				node = e.On[0]
+			}
+		}
+	}
+	if node == "" {
+		return fmt.Errorf("%s is assigned nowhere; `assign <machine> %s` first, or say --node", *module, *module)
+	}
+	address, err := broker.BusAddress()
+	if err != nil {
+		return err
+	}
+	return issueOnTheNewBus(ctx, inv, m, node, address)
+}
+
+// buildBehind builds every module the mesh holds older than its source has.
+//
+// **This is the loop novox/hq ADR 0010 replaced a pipeline with, closed.** The mesh already
+// records where each module came from and what its source last had; until this, a person read
+// that list and retyped each repository — which is a person being the loop, and the thing a
+// pipeline was doing before it was taken away.
+//
+// Each is built and recorded on its own. **One failing does not stop the others**, for the same
+// reason one broken module no longer blocks a machine's whole declaration: a mesh where one bad
+// repository holds back nine good ones is a mesh where nobody dares add the tenth.
+func buildBehind(ctx context.Context, wait time.Duration) error {
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	held, err := inv.Catalogued(ctx)
+	if err != nil {
+		return err
+	}
+	var stale []inventory.Entry
+	for _, e := range held {
+		if !e.Source.Current() {
+			stale = append(stale, e)
+		}
+	}
+	if len(stale) == 0 {
+		// Said rather than doing nothing quietly: "nothing needed building" and "this did not
+		// run" must never look the same.
+		fmt.Println("every module the mesh holds is what its source last had")
+		return nil
+	}
+
+	fmt.Printf("%d module(s) behind their source:\n", len(stale))
+	for _, e := range stale {
+		fmt.Printf("  %s  %s < %s\n",
+			e.Manifest.Module, short(e.Source.BuiltFrom), short(e.Source.Head))
+	}
+	fmt.Println()
+
+	// Bases first: a module built before the module it stands on is built against the old one
+	// and reports success (novox/hq 04-ISSUES/131).
+	against, err := inv.BuiltAgainst(ctx)
+	if err != nil {
+		return err
+	}
+	stale = orderByBases(stale, against)
+
+	var failed []string
+	for _, e := range stale {
+		fmt.Printf("--- %s\n", e.Manifest.Module)
+		// Its own recorded ref, not its head commit: a module tracking a branch should be built
+		// from that branch, and pinning to the commit the mesh happened to notice would quietly
+		// turn a tracked branch into a pin.
+		source := buildSource{Repository: e.Source.Repository, Seat: e.Source.Seat}
+		if err := buildOne(ctx, source, e.Source.Path, e.Source.Ref, wait); err != nil {
+			fmt.Printf("  %v\n", err)
+			failed = append(failed, e.Manifest.Module)
+		}
+	}
+
+	if len(failed) > 0 {
+		return fmt.Errorf("%d of %d could not be built: %s",
+			len(failed), len(stale), strings.Join(failed, ", "))
+	}
+	fmt.Printf("\n%d module(s) built. `push --behind` sends them to the machines running them\n",
+		len(stale))
+	return nil
+}
+
+// buildOne asks a build machine for one repository and records everything that came back.
+//
+// Separated from the command so `--behind` can walk a list without a second path to the same act.
+func buildOne(ctx context.Context, source buildSource, path, ref string, wait time.Duration) error {
+	_, err := buildOneAsked(ctx, source, path, ref, wait, false)
+	return err
+}
+
+// buildOneAsked is buildOne answering the id it asked with — what a plan keeps to match the outcome
+// by (novox/hq ADR 0219) — and, for an ask not waited for, optionally a dry run: built and looked
+// at, never taken in (issue 240), which is what `replay` asks unless told to register.
+func buildOneAsked(ctx context.Context, source buildSource, path, ref string, wait time.Duration,
+	dryRun bool) (string, error) {
+	if dryRun && wait != 0 {
+		return "", errors.New("a dry run waited for is `build --dry-run`")
+	}
+	// Before anything is asked of a builder: a source on a seat nobody holds is refused here, with
+	// the reason, rather than sent to a machine to fail at `git clone`.
+	repository, err := cloneFrom(ctx, source)
+	if err != nil {
+		return "", err
+	}
+
+	ident, err := openIdentity(ctx)
+	if err != nil {
+		return "", err
+	}
+	defer ident.Close()
+
+	server, err := connectLink(ctx, nil, nil, nil)
+	if err != nil {
+		return "", err
+	}
+	defer server.Close()
+
+	// Correlated by something the control plane makes, not by the module's name: two builds of one
+	// module can be in flight, and the second answer is not the first one's.
+	request := link.BuildRequest{
+		ID:         link.NewBuildID(time.Now()),
+		Repository: repository,
+		Path:       path,
+		Ref:        ref,
+		Held:       heldBy(ctx),
+		Seats:      seatBases(ctx),
+		DryRun:     dryRun,
+	}
+	fmt.Printf("asked for %s", source)
+	if source.Seat != "" {
+		fmt.Printf(" (%s)", repository)
+	}
+	// The id is how a person follows this build while it runs: `builds --log <id>`.
+	fmt.Printf(" as %s", request.ID)
+	if path != "" {
+		fmt.Printf(" at %s", path)
+	}
+	if ref != "" {
+		fmt.Printf(" on %s", ref)
+	}
+	fmt.Println()
+
+	seat := buildSeatHeld(ctx)
+	ask, err := askOverOn(seat)
+	if err != nil {
+		return "", err
+	}
+	defer ask.Close()
+	fmt.Printf("  of %s\n", seat)
+
+	if wait == 0 {
+		// Asked and not waited for (novox/hq issue 176): the outcome is the role's event, and the
+		// controller takes it in — records the build, registers the module — whether or not anybody
+		// is still here. A tool call cannot hold a connection for the minutes a build takes; it
+		// follows the build by its id instead.
+		if err := ask.Ask(ctx, request); err != nil {
+			return "", err
+		}
+		if dryRun {
+			fmt.Printf("asked as a dry run, not waited for: `builds --log %s` follows it as it runs; "+
+				"its outcome is not taken in\n", request.ID)
+			return request.ID, nil
+		}
+		fmt.Printf("asked, not waited for: `builds --log %s` follows it as it runs, and `builds` "+
+			"shows what came of it; the module is registered when the outcome comes\n", request.ID)
+		return request.ID, nil
+	}
+
+	result, err := ask.Submit(ctx, request, wait)
+	if err != nil {
+		return request.ID, err
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return request.ID, err
+	}
+	defer open.Close()
+	manifest, kept, err := takeIn(ctx, open.inventory, result)
+	if err != nil {
+		return request.ID, err
+	}
+	// Said as recorded: what each artifact is, not where this builder happened to push it.
+	for _, made := range kept.Made {
+		fmt.Printf("  %-12s %s %s\n", made.Name, made.Kind, made.Reference)
+	}
+	fmt.Printf("\n%s %s, built on %s from %s\n",
+		manifest.Module, manifest.Version, result.On, short(result.Commit))
+	saysWhenThePolicyActs(ctx, open.inventory, manifest.Module)
+	fmt.Printf("  run `assign <node> %s` to put it somewhere\n", manifest.Module)
+	return request.ID, nil
+}
+
+// saysWhenThePolicyActs tells whoever built a module that its upgrade policy will send the
+// result on at once (novox/hq issue 126, ADR 0163): a person choreographing a data move must
+// know which module will not wait for them.
+func saysWhenThePolicyActs(ctx context.Context, inv *inventory.Inventory, module string) {
+	if u, err := inv.UpgradeOf(ctx, module); err == nil && u.RollOut {
+		how := "one machine at a time"
+		if u.Together {
+			how = "every machine at once"
+		}
+		fmt.Printf("  %s rolls out on build: the machines running it are sent this now, %s — "+
+			"`upgrade %s record` first if something must move before it does\n", module, how, module)
+	}
+}
+
+// takeIn is what the mesh does with a build's outcome, whoever hears it: the waiting command and
+// the daemon that follows the role's events both come here (novox/hq issue 176), so a build's
+// result reaches the catalogue whether or not the asker was still listening.
+//
+// Kept before it is judged. A failed build that leaves no trace is indistinguishable from one
+// nobody asked for, and the difference is the whole of whether somebody should be looking at
+// something. Then parsed with the same parser a hand-written manifest goes through — a second path
+// would be a second thing to disagree about what a manifest is — and registered with where it came
+// from: **for a source on a seat, as the path and the seat, never the URL just cloned** (ADR 0111),
+// which the request carried and the outcome echoes. A definition naming an installation is refused
+// here, where it would enter the catalogue; the build stays recorded and the refusal says which.
+//
+// Idempotent: the same outcome taken in twice registers the same module twice, which is one row
+// written with the same values.
+func takeIn(ctx context.Context, inv *inventory.Inventory, result link.BuildResult) (
+	catalogue.Manifest, inventory.Build, error) {
+	kept := buildFrom(result)
+	if err := inv.RecordBuild(ctx, kept); err != nil {
+		return catalogue.Manifest{}, kept, err
+	}
+	if result.Failed != "" {
+		// The builder's own words. Wrapping them in something about the control plane would put
+		// two explanations between a person and a build log.
+		return catalogue.Manifest{}, kept, fmt.Errorf("%s could not build %s:\n%s",
+			result.On, result.Repository, result.Failed)
+	}
+	manifest, err := catalogue.ParseManifest(kept.Manifest)
+	if err != nil {
+		return catalogue.Manifest{}, kept, fmt.Errorf("%s built %s and what came back is not a manifest: %w",
+			result.On, result.Repository, err)
+	}
+	recorded := inventory.Source{
+		Repository: result.Repository, Path: result.Path, Ref: result.Ref,
+		BuiltFrom: result.Commit, Head: result.Commit,
+		// What it stood on, so registration can judge a built manifest's base (to-be 38 WP2.4).
+		Against: kept.Against,
+		// When it was asked, so an older request heard later does not replace a newer one
+		// (novox/hq 04-ISSUES/219).
+		Asked: kept.Asked,
+	}
+	if result.Source != nil && result.Source.Seat != "" {
+		recorded.Repository, recorded.Seat = result.Source.Repository, result.Source.Seat
+	}
+	// **A build at a commit does not change the branch a module follows** (novox/hq 04-ISSUES/215):
+	// the commit is built and recorded as what it was built from, and the module keeps following
+	// what it followed before — the repository's default branch for one new to the catalogue.
+	if followedBranch(result.Ref) == "" && result.Ref != "" {
+		recorded.Ref = ""
+		if was, err := inv.SourceOf(ctx, manifest.Module); err == nil {
+			recorded.Ref = followedBranch(was.Ref)
+		}
+	}
+	if err := namesNoInstallation(manifest); err != nil {
+		return manifest, kept, fmt.Errorf("%s built %s (%s), and the mesh does not register it: %w",
+			result.On, result.Repository, short(result.Commit), err)
+	}
+	if err := inv.RegisterModule(ctx, manifest, recorded); err != nil {
+		if errors.Is(err, inventory.ErrSuperseded) {
+			return manifest, kept, fmt.Errorf("%s built %s (%s), recorded and not registered: %w",
+				result.On, manifest.Module, short(result.Commit), err)
+		}
+		return manifest, kept, err
+	}
+	// The keep set just moved, and new bytes just landed (novox/hq ADR 0189). Asked here rather
+	// than on a timer of its own: this is the only moment either is true. Never fatal — the build
+	// worked and the module is registered.
+	collect(ctx, inv)
+	return manifest, kept, nil
+}
+
+// buildAndShow builds and prints the manifest without recording anything.
+func buildAndShow(ctx context.Context, source buildSource, path, ref string, wait time.Duration) error {
+	repository, err := cloneFrom(ctx, source)
+	if err != nil {
+		return err
+	}
+	ident, err := openIdentity(ctx)
+	if err != nil {
+		return err
+	}
+	defer ident.Close()
+	server, err := connectLink(ctx, nil, nil, nil)
+	if err != nil {
+		return err
+	}
+	defer server.Close()
+
+	ask, err := askOverOn(buildSeatHeld(ctx))
+	if err != nil {
+		return err
+	}
+	defer ask.Close()
+
+	result, err := ask.Submit(ctx, link.BuildRequest{
+		ID:         link.NewBuildID(time.Now()),
+		Repository: repository, Path: path, Ref: ref,
+		Held: heldBy(ctx), Seats: seatBases(ctx),
+		DryRun: true,
+	}, wait)
+	if err != nil {
+		return err
+	}
+	if result.Failed != "" {
+		return fmt.Errorf("%s could not build %s:\n%s", result.On, result.Repository, result.Failed)
+	}
+	manifest, err := catalogue.ParseManifest(result.Manifest)
+	if err != nil {
+		return fmt.Errorf("%s built %s and what came back is not a manifest: %w",
+			result.On, result.Repository, err)
+	}
+	body, err := json.MarshalIndent(manifest, "", "  ")
+	if err != nil {
+		return err
+	}
+	fmt.Println(string(body))
+	return nil
+}
+
+// answers is what the three questions came back with, read once.
+type answers struct {
+	wrong   []inventory.Doing
+	nodes   []inventory.Node
+	quiet   []inventory.Node
+	behind  map[string][]string
+	sources map[string]inventory.Source
+	// waiting is every machine not running what the mesh would send it.
+	waiting []inventory.Machine
+	// reported is every machine's last word beside when it was last sent a declaration — the
+	// pair that answers "has it caught up", which waiting alone cannot (the sent digest is
+	// recorded at send, not at apply).
+	reported []inventory.Reported
+	// plans is what the last merges produced and where each stands (novox/hq ADR 0162).
+	plans []inventory.Plan
+	// paused is whether the build seat takes work, which a plan waiting on it says (ADR 0219).
+	paused pauseView
+	// refused is why a machine cannot be worked out at all, by name. A different thing from every
+	// other answer here: those are about a machine that was told something, and this is about one
+	// that cannot be told anything — it never reaches waiting, because nothing was computed for it
+	// to compare against, so without this a wholly blocked mesh reads as a well one.
+	refused map[string]string
+	// network is why the private network could not be computed, when it could not. Almost always
+	// a consequence of the refusals above: a node that does not resolve is not on the network, and
+	// a mesh whose hub is that node has no hub.
+	network string
+	// filtered is every converged machine that is not filtered by the mesh alone (novox/hq ADR
+	// 0168): what filters it beyond the mesh's own, the runtime's plumbing and bans, by name — a
+	// predecessor's chain, a found firewall in force again. Such a machine is not "all well".
+	filtered map[string]inventory.Filtering
+	// untaken is, per machine, each assigned module whose resources the machine is holding as it
+	// found them, and how many — a module that was assigned, sent, and is running none of what it
+	// declares because nothing has taken it (novox/hq ADR 0100, 04-ISSUES/125).
+	//
+	// **Its absence cost an outage.** The module was assigned, the push reported success, this
+	// command said the machine was doing everything it was told, and the module's three containers
+	// did not exist. On the strength of those reports the predecessor's proxy was stopped and every
+	// public name on the machine went dark. The holds were correct; they were recorded only in the
+	// machine's own state file, and the one visible symptom was a count that did not add up.
+	untaken map[string]map[string]int
+	// unheld is every module on a machine whose resources are applied through a seat nothing on
+	// that machine holds (novox/hq ADR 0207), with the modules that could hold it. Reported, not
+	// refused, until the switch — and while there is any, the mesh is not all well: the order the
+	// machines' modules are built in is the mesh's to keep, and this is where it says it is not kept.
+	unheld []catalogue.Unheld
+	// failing is every consumer a provider says it keeps failing (novox/hq ADR 0224): a provider's
+	// journal was the only place that said so for a day (04-ISSUES/179).
+	failing []inventory.ProviderStanding
+}
+
+// heldBy is every artifact this mesh has built, for a build that may need one as its base.
+//
+// **A failure here is not a failure to build.** A module that names no base does not need this at
+// all, and one that does gets a refusal naming exactly what is missing — which is a better sentence
+// than a build command refusing to start because a query did not run. So the store not opening is
+// reported and the build goes ahead without it.
+//
+// Routed through the artifact store as the network reaches it now (novox/hq 04-ISSUES/102): a
+// base is recorded by digest and path, and a build machine needs something it can pull.
+func heldBy(ctx context.Context) map[string]string {
+	open, err := openStores(ctx)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not read what this mesh has built, so a module naming a "+
+			"base will be told that base is missing: %v\n", err)
+		return nil
+	}
+	defer open.Close()
+	held, err := open.inventory.Held(ctx)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not read what this mesh has built: %v\n", err)
+		return nil
+	}
+	address, err := whereABuilderReachesTheStore(ctx, open.inventory)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not find the artifact store on this mesh's network, so a "+
+			"module naming a base will be handed a reference nothing can fetch: %v\n", err)
+		return held
+	}
+	if address == "" {
+		return held
+	}
+	routed := make(map[string]string, len(held))
+	for repository, reference := range held {
+		routed[repository] = catalogue.Rerouted(reference, address)
+	}
+	return routed
+}
+
+// askOver opens the way a build is asked for, on whichever bus the mesh is on.
+//
+// **One place chooses**, as everywhere else the bus change went (novox/hq ADR 0116 step 5). On the bus
+// the mesh runs on today this needs the controller's own connection, so it is handed one; on the bus
+// being built it dials, because a build request is a one-shot and holds nothing else.
+func askOverOn(seat string) (link.Builders, error) {
+	address, err := broker.BusAddress()
+	if err != nil {
+		return nil, err
+	}
+	return link.BuildsOverNATSOn(address, seat)
+}
+
+// buildSeatHeld is the build role to ask: the one some assigned module claims (novox/hq ADR 0190,
+// the handover). Read from the catalogue at ask time, because the answer changes exactly once, the
+// moment the first build-agent is assigned — and a controller that asked the new role before then
+// would queue work nothing takes, while the outcome that registers build-agent itself has to come
+// from the old builder. When the catalogue cannot be read the current role is asked, said aloud.
+func buildSeatHeld(ctx context.Context) string {
+	open, err := openStores(ctx)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not read what is assigned, so the build is asked of %s: %v\n",
+			link.TheBuildMachine, err)
+		return link.TheBuildMachine
+	}
+	defer open.Close()
+	entries, err := open.inventory.Catalogued(ctx)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not read the catalogue, so the build is asked of %s: %v\n",
+			link.TheBuildMachine, err)
+		return link.TheBuildMachine
+	}
+	return buildSeatAmong(entries)
+}
+
+// buildSeatAmong is the rule, over what the catalogue holds: the current build role when any
+// assigned module claims it; else the retired role while an assigned module still claims that; else
+// the current role, which is where every ask goes once the handover is done.
+func buildSeatAmong(entries []inventory.Entry) string {
+	heldBefore := false
+	for _, e := range entries {
+		if len(e.On) == 0 {
+			continue
+		}
+		if e.Manifest.ClaimsSeat(link.TheBuildMachine) {
+			return link.TheBuildMachine
+		}
+		if e.Manifest.ClaimsSeat(link.TheBuildMachineBefore) {
+			heldBefore = true
+		}
+	}
+	if heldBefore {
+		return link.TheBuildMachineBefore
+	}
+	return link.TheBuildMachine
+}
+
+// buildLog prints everything a build machine said about one build, read back from the bus.
+//
+// **From the stream, not from a record** (novox/hq ADR 0157). A build's lines are the role's own
+// events under the build's id, retained with every other event; the mesh keeps no second copy. Read
+// with a consumer of its own that is gone when this returns, so nothing accumulates in the server
+// for the reading, and filtered by subject, so one build's lines are all that travel.
+func buildLog(ctx context.Context, id string) error {
+	address, err := broker.BusAddress()
+	if err != nil {
+		return err
+	}
+	js, err := broker.Dial(address)
+	if err != nil {
+		return fmt.Errorf("cannot reach the bus to read a build's log: %w", err)
+	}
+	defer js.Close()
+
+	// Under whichever build role did it: a build asked of the retired role during the handover
+	// (ADR 0190) said its lines as that role's events, and a reader should not have to know which.
+	lines := link.BuildLogOf("*", id)
+	sub, err := js.Context().PullSubscribe(lines, "",
+		nats.BindStream(broker.EventsStream), nats.DeliverAll(), nats.AckNone())
+	if err != nil {
+		return fmt.Errorf("cannot read %s from the bus: %w", lines, err)
+	}
+	defer func() { _ = sub.Unsubscribe() }()
+
+	printed := 0
+	for {
+		batch, err := sub.Fetch(200, nats.MaxWait(2*time.Second))
+		if err != nil && !errors.Is(err, nats.ErrTimeout) && !errors.Is(err, context.DeadlineExceeded) {
+			return fmt.Errorf("reading a build's log: %w", err)
+		}
+		for _, msg := range batch {
+			var line link.BuildLine
+			if err := json.Unmarshal(msg.Data, &line); err != nil {
+				fmt.Printf("  ? %s\n", string(msg.Data))
+				continue
+			}
+			at := line.At
+			if t, err := time.Parse(time.RFC3339Nano, line.At); err == nil {
+				at = t.Local().Format("15:04:05")
+			}
+			fmt.Printf("%s %4d [%s] %s\n", at, line.Seq, line.Step, line.Message)
+			printed++
+		}
+		if len(batch) < 200 {
+			break
+		}
+	}
+	if printed == 0 {
+		fmt.Printf("nothing on the bus for build %s: no build by that id in the last week, or a build "+
+			"machine older than this that said nothing while building\n", id)
+	}
+	return nil
+}

cmd/mesh-controller/build_seat_test.go

@@ -0,0 +1,43 @@
+package main
+
+import (
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+func claiming(module, seat string, on ...string) inventory.Entry {
+	return inventory.Entry{
+		Manifest: catalogue.Manifest{Module: module, Claims: []catalogue.Claim{{Name: seat}}},
+		On:       on,
+	}
+}
+
+// The controller asks the build role that has a holder (novox/hq ADR 0190 handover): the retired
+// one while only the builder is assigned, the current one from the first build-agent on, and the
+// current one when nothing holds either — where every ask goes once the handover is done.
+func TestTheControllerAsksTheBuildRoleThatHasAHolder(t *testing.T) {
+	onlyTheBuilder := []inventory.Entry{
+		claiming("builder", link.TheBuildMachineBefore, "anchor"),
+		claiming("build-agent", link.TheBuildMachine), // registered, assigned nowhere yet
+	}
+	if got := buildSeatAmong(onlyTheBuilder); got != link.TheBuildMachineBefore {
+		t.Errorf("with only the builder assigned, asked %q", got)
+	}
+	bothHeld := []inventory.Entry{
+		claiming("builder", link.TheBuildMachineBefore, "anchor"),
+		claiming("build-agent", link.TheBuildMachine, "home-server"),
+	}
+	if got := buildSeatAmong(bothHeld); got != link.TheBuildMachine {
+		t.Errorf("with a build-agent assigned anywhere, asked %q", got)
+	}
+	neither := []inventory.Entry{claiming("builder", link.TheBuildMachineBefore)}
+	if got := buildSeatAmong(neither); got != link.TheBuildMachine {
+		t.Errorf("with no holder of either, asked %q, want the current role", got)
+	}
+	if got := buildSeatAmong(nil); got != link.TheBuildMachine {
+		t.Errorf("an empty catalogue asks %q", got)
+	}
+}

cmd/mesh-controller/build_test.go

@@ -0,0 +1,152 @@
+package main
+
+import (
+	"encoding/json"
+	"errors"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A build's outcome is taken in the same way whoever hears it (novox/hq issue 176): recorded, and
+// the module registered with its source as the seat and path when the request said so — never the
+// URL. A definition naming an installation is recorded and not registered; a failure is recorded
+// and said.
+func TestABuildHeardIsRecordedAndRegistered(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	manifest, _ := json.Marshal(map[string]any{"module": "shop", "version": "3"})
+	m, _, err := takeIn(ctx, open.inventory, link.BuildResult{
+		ID: "b-1", Repository: "http://forge.internal:20000/novox/shop.git", Path: "modules/shop",
+		Ref: "main", On: "anchor", Commit: "abcdef0123", Manifest: manifest,
+		Source: &link.SourceOnSeat{Seat: "git", Repository: "novox/shop"},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if m.Module != "shop" {
+		t.Fatalf("registered %q", m.Module)
+	}
+	shelf, err := open.inventory.Catalogue(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, held := shelf["shop"]; !held {
+		t.Fatal("the module a heard build produced is not in the catalogue")
+	}
+	src, err := open.inventory.SourceOf(ctx, "shop")
+	if err != nil || src.Seat != "git" || src.Repository != "novox/shop" || src.BuiltFrom != "abcdef0123" {
+		t.Fatalf("the source is the seat and the path, never the URL: %+v %v", src, err)
+	}
+	builds, err := open.inventory.Builds(ctx, "shop", 5)
+	if err != nil || len(builds) != 1 || builds[0].ID != "b-1" {
+		t.Fatalf("the build is not recorded once: %v %v", builds, err)
+	}
+
+	named, _ := json.Marshal(map[string]any{"module": "idp", "version": "1", "resources": []any{
+		map[string]any{"id": "server", "type": "container", "image": "x@sha256:aa",
+			"env": map[string]any{"KC_HOSTNAME": "https://login.mesh-one.be"}}}})
+	_, _, err = takeIn(ctx, open.inventory, link.BuildResult{
+		ID: "b-2", Repository: "/r", On: "anchor", Commit: "0123456789", Manifest: named})
+	if err == nil || !strings.Contains(err.Error(), "does not register it") {
+		t.Fatalf("a definition naming an installation was taken in: %v", err)
+	}
+	if shelf, _ := open.inventory.Catalogue(ctx); shelf["idp"].Module != "" {
+		t.Fatal("the refused module was registered anyway")
+	}
+	if builds, _ := open.inventory.Builds(ctx, "idp", 5); len(builds) != 1 {
+		t.Fatalf("the refused build was not recorded: %v", builds)
+	}
+
+	_, _, err = takeIn(ctx, open.inventory, link.BuildResult{ID: "b-3", Repository: "/r", On: "anchor", Failed: "no compiler"})
+	if err == nil || !strings.Contains(err.Error(), "no compiler") {
+		t.Fatalf("a failure is said in the builder's words: %v", err)
+	}
+}
+
+// novox/hq 04-ISSUES/215: a build asked at a commit is recorded as built from that commit, and the
+// module keeps following the branch it followed — a new one, the default branch.
+func TestABuildAtACommitKeepsTheBranchTheModuleFollows(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	manifest, _ := json.Marshal(map[string]any{"module": "unifi", "version": "1"})
+	result := func(id, ref, commit string) link.BuildResult {
+		return link.BuildResult{ID: id, Repository: "http://forge.internal:20000/novox/mesh-catalog.git",
+			Path: "modules/unifi", Ref: ref, On: "anchor", Commit: commit, Manifest: manifest,
+			Source: &link.SourceOnSeat{Seat: "git", Repository: "novox/mesh-catalog"}}
+	}
+	if _, _, err := takeIn(ctx, open.inventory, result("b-1", "main", "1111111aaaa")); err != nil {
+		t.Fatal(err)
+	}
+	if _, _, err := takeIn(ctx, open.inventory, result("b-2", "9c97a8a", "9c97a8a1d2c3")); err != nil {
+		t.Fatal(err)
+	}
+	src, err := open.inventory.SourceOf(ctx, "unifi")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if src.Ref != "main" || src.BuiltFrom != "9c97a8a1d2c3" {
+		t.Errorf("after a build at a commit the module follows %q, built from %q; want main, 9c97a8a1d2c3", src.Ref, src.BuiltFrom)
+	}
+
+	// One new to the catalogue, first built at a commit, follows the default branch.
+	other, _ := json.Marshal(map[string]any{"module": "letta", "version": "1"})
+	r := result("b-3", "deadbeef", "deadbeefcafe")
+	r.Manifest, r.Path = other, "modules/letta"
+	if _, _, err := takeIn(ctx, open.inventory, r); err != nil {
+		t.Fatal(err)
+	}
+	if src, _ := open.inventory.SourceOf(ctx, "letta"); src.Ref != "" {
+		t.Errorf("a module first built at a commit follows %q, want the default branch", src.Ref)
+	}
+}
+
+// novox/hq 04-ISSUES/219: an older request heard after a newer one is recorded and not registered,
+// so a push sends what the newer request built.
+func TestAnOlderBuildHeardLaterDoesNotReplaceTheNewer(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	older := time.Date(2026, 10, 3, 21, 33, 45, 0, time.UTC)
+	newer := time.Date(2026, 10, 3, 21, 51, 57, 0, time.UTC)
+	result := func(asked time.Time, image string) link.BuildResult {
+		manifest, _ := json.Marshal(map[string]any{"module": "postgres", "version": image})
+		return link.BuildResult{ID: link.NewBuildID(asked), Repository: "http://forge.internal:20000/novox/mesh-catalog.git",
+			Path: "modules/postgres", Ref: "main", On: "anchor", Commit: "efff5415", Manifest: manifest,
+			Source: &link.SourceOnSeat{Seat: "git", Repository: "novox/mesh-catalog"}}
+	}
+	if _, _, err := takeIn(ctx, open.inventory, result(newer, "4bcd5f73")); err != nil {
+		t.Fatal(err)
+	}
+	_, _, err := takeIn(ctx, open.inventory, result(older, "0ab07fa9"))
+	if !errors.Is(err, inventory.ErrSuperseded) {
+		t.Fatalf("the older request's outcome was taken in as current: %v", err)
+	}
+	shelf, err := open.inventory.Catalogue(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := shelf["postgres"].Version; got != "4bcd5f73" {
+		t.Errorf("postgres is %q; want the newer request's 4bcd5f73", got)
+	}
+	if builds, _ := open.inventory.Builds(ctx, "postgres", 5); len(builds) != 2 {
+		t.Errorf("the late build was not recorded: %v", builds)
+	}
+}
+
+func TestABuildIDSaysWhenItWasAsked(t *testing.T) {
+	at := time.Date(2026, 10, 3, 21, 51, 57, 392539762, time.UTC)
+	if got, ok := link.BuildAskedAt(link.NewBuildID(at)); !ok || !got.Equal(at) {
+		t.Errorf("read back %v %v; want %v", got, ok, at)
+	}
+	if got, ok := link.BuildAskedAt("build-1791064317392539762"); !ok || got.Format(time.TimeOnly) != "21:51:57" {
+		t.Errorf("the incident's id reads as %v %v", got, ok)
+	}
+	for _, id := range []string{"b-1", "build-2", "build-", "build-x", ""} {
+		if _, ok := link.BuildAskedAt(id); ok {
+			t.Errorf("%q read as a request time", id)
+		}
+	}
+}

cmd/mesh-controller/buscertificate.go

@@ -0,0 +1,258 @@
+package main
+
+import (
+	"context"
+	"crypto/rand"
+	"crypto/rsa"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/pem"
+	"errors"
+	"fmt"
+	"math/big"
+	"net"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// The bus's own certificate, made by the mesh rather than borrowed from an image.
+//
+// **The foundation asked a third-party image for a tool it never said must be there** (novox/hq
+// 04-ISSUES/146). The bootstrap made this certificate by running `openssl` inside the broker's
+// image, which worked while the broker was one that happened to carry it and stopped the day the
+// bus changed: the new one has a shell and no openssl, so the step exited 127 and no mesh could be
+// raised. Substituting another image the bundle names does not help — none of them carry it
+// either.
+//
+// So the program that needs a certificate makes one. It is the mesh's own binary, already on the
+// machine at this point in the bootstrap (the schema step ran it), and it needs nothing from the
+// image it writes into but a mounted directory.
+//
+// **Self-signed, and that is the design** — a host pins this server's exact certificate and
+// authenticates with a password (novox/hq ADR 0004). There is no authority above it to ask, and at
+// this moment in a bootstrap there is no mesh to ask one of.
+//
+// Idempotent, because the step is applied again on every reconcile and a second certificate would
+// be one the hosts that pinned the first no longer believe.
+
+// busCertificateNames is what the bus is reached by: the container name on a mesh network, and the
+// loopback address the machine's own foundation dials.
+var busCertificateNames = []string{"mesh-broker"}
+
+const busCertificateLife = 10 * 365 * 24 * time.Hour
+
+// busCertificate makes the bus's certificate in a directory, or says whether one is there.
+//
+//	broker certificate --into /tls           make it if it is not there
+//	broker certificate --check --into /tls   exit non-zero unless a usable pair is
+func busCertificate(args []string) error {
+	into, check := "", false
+	for i := 0; i < len(args); i++ {
+		switch args[i] {
+		case "--check":
+			check = true
+		case "--into":
+			if i+1 >= len(args) {
+				return errors.New("--into needs a directory")
+			}
+			into = args[i+1]
+			i++
+		default:
+			return fmt.Errorf("broker certificate [--check] --into <directory>: %q", args[i])
+		}
+	}
+	if into == "" {
+		return errors.New("broker certificate [--check] --into <directory>")
+	}
+	crt, key := filepath.Join(into, "tls.crt"), filepath.Join(into, "tls.key")
+
+	if usable, err := busCertificateUsable(crt, key); err != nil {
+		return err
+	} else if usable {
+		fmt.Printf("the bus already has a certificate at %s, and it was left alone\n", crt)
+		return nil
+	}
+	if check {
+		// Said as a failure, because that is what the caller asked: a bootstrap's verify runs
+		// this and a false answer is what makes the step run.
+		return fmt.Errorf("no usable certificate and key at %s", into)
+	}
+	return writeBusCertificate(crt, key)
+}
+
+// busCertificateUsable says whether a certificate and its key are both there and parse.
+//
+// Both, and parsed rather than stat'ed: a half-written pair is the state a bootstrap interrupted
+// between the two files leaves behind, and a step that treated it as done would hand the server a
+// certificate with no key and report success.
+func busCertificateUsable(crt, key string) (bool, error) {
+	certPEM, err := os.ReadFile(crt)
+	if errors.Is(err, os.ErrNotExist) {
+		return false, nil
+	}
+	if err != nil {
+		return false, err
+	}
+	keyPEM, err := os.ReadFile(key)
+	if errors.Is(err, os.ErrNotExist) {
+		return false, nil
+	}
+	if err != nil {
+		return false, err
+	}
+	if _, err := tlsPairParses(certPEM, keyPEM); err != nil {
+		return false, nil
+	}
+	return true, nil
+}
+
+func tlsPairParses(certPEM, keyPEM []byte) (*x509.Certificate, error) {
+	block, _ := pem.Decode(certPEM)
+	if block == nil || block.Type != "CERTIFICATE" {
+		return nil, errors.New("not a certificate")
+	}
+	certificate, err := x509.ParseCertificate(block.Bytes)
+	if err != nil {
+		return nil, err
+	}
+	keyBlock, _ := pem.Decode(keyPEM)
+	if keyBlock == nil {
+		return nil, errors.New("not a key")
+	}
+	if _, err := x509.ParsePKCS8PrivateKey(keyBlock.Bytes); err != nil {
+		if _, err := x509.ParsePKCS1PrivateKey(keyBlock.Bytes); err != nil {
+			return nil, err
+		}
+	}
+	return certificate, nil
+}
+
+func writeBusCertificate(crt, key string) error {
+	private, err := rsa.GenerateKey(rand.Reader, 2048)
+	if err != nil {
+		return err
+	}
+	serial, err := rand.Int(rand.Reader, new(big.Int).Lsh(big.NewInt(1), 128))
+	if err != nil {
+		return err
+	}
+	template := &x509.Certificate{
+		SerialNumber:          serial,
+		Subject:               pkix.Name{CommonName: busCertificateNames[0]},
+		DNSNames:              busCertificateNames,
+		IPAddresses:           []net.IP{net.ParseIP("127.0.0.1")},
+		NotBefore:             time.Now().Add(-time.Hour),
+		NotAfter:              time.Now().Add(busCertificateLife),
+		KeyUsage:              x509.KeyUsageDigitalSignature | x509.KeyUsageKeyEncipherment,
+		ExtKeyUsage:           []x509.ExtKeyUsage{x509.ExtKeyUsageServerAuth},
+		BasicConstraintsValid: true,
+	}
+	der, err := x509.CreateCertificate(rand.Reader, template, template, &private.PublicKey, private)
+	if err != nil {
+		return err
+	}
+	pkcs8, err := x509.MarshalPKCS8PrivateKey(private)
+	if err != nil {
+		return err
+	}
+
+	// **The key first, and only then the certificate**, so the pair a reader finds is never a
+	// certificate whose key has not been written yet — the one order in which an interruption
+	// leaves something that looks finished (novox/hq 04-ISSUES/014, a key present and unusable).
+	if err := os.WriteFile(key, pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: pkcs8}), 0o600); err != nil {
+		return err
+	}
+	if err := os.WriteFile(crt, pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: der}), 0o644); err != nil {
+		return err
+	}
+	fmt.Printf("made the bus a certificate for %v, valid until %s\n  %s\n  %s\n",
+		busCertificateNames, template.NotAfter.Format(time.RFC3339), crt, key)
+	return nil
+}
+
+// busAccounts writes the mesh's composed user list to a file.
+//
+// **For genesis, where no declaration can deliver it** (novox/hq 04-ISSUES/146). Everywhere else
+// the list reaches the machine running the bus as a resource of the module that holds it — which
+// requires that machine to be an enrolled node, and at genesis it is not: the first node cannot
+// enrol because the account it would enrol with cannot be composed onto a bus it has no declaration
+// for. The installer breaks that circle by placing the file itself, once, and the module takes the
+// file over from its first push.
+//
+// The same composition, not a second one: this asks the store for the same records and renders them
+// with the same composer the declaration uses. A genesis that hand-wrote an account would be a
+// second statement of who may say what, able to disagree with the first.
+//
+// **It writes to standard output unless told a file**, and that is the point: the control plane
+// composes and says what it composed, and whoever is raising the machine puts it where that
+// machine's bus reads it. A control plane that wrote into the bus's own directory would have to
+// know where that is and how to make the server re-read it — which is the module's knowledge, and
+// the module is what takes this over on the first push.
+//
+//	broker accounts > /var/lib/mesh-bus-conf/accounts.conf
+func busAccounts(ctx context.Context, args []string) error {
+	into := ""
+	for i := 0; i < len(args); i++ {
+		switch args[i] {
+		case "--into":
+			if i+1 >= len(args) {
+				return errors.New("--into needs a file")
+			}
+			into = args[i+1]
+			i++
+		default:
+			return fmt.Errorf("broker accounts --into <file>: %q", args[i])
+		}
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+
+	records, err := open.inventory.BusRecords(ctx)
+	if err != nil {
+		return err
+	}
+	users, err := broker.Users(records)
+	if err != nil {
+		return err
+	}
+	kept, err := open.inventory.BusUsers(ctx)
+	if err != nil {
+		return err
+	}
+	hashes := make(map[string]string, len(kept))
+	for name, u := range kept {
+		hashes[name] = u.PasswordHash
+	}
+	filled, missing := broker.WithPasswords(users, hashes)
+	if len(missing) > 0 {
+		// To standard error, always: the composed file may be going to standard output, and a
+		// remark in the middle of it is a configuration the server refuses to parse.
+		fmt.Fprintf(os.Stderr, "leaving out %d user(s) the mesh has minted no credential for: %s\n",
+			len(missing), strings.Join(missing, ", "))
+	}
+	if len(filled) == 0 {
+		return errors.New("not one user has a credential, so this list would refuse every " +
+			"connection in the mesh")
+	}
+	accounts, err := broker.ComposeAccounts(filled)
+	if err != nil {
+		return err
+	}
+	if into == "" {
+		fmt.Print(accounts)
+		return nil
+	}
+	if err := os.WriteFile(into, []byte(accounts), 0o600); err != nil {
+		return err
+	}
+	fmt.Printf("wrote %d user(s) to %s\n", len(filled), into)
+	return nil
+}

cmd/mesh-controller/buscertificate_test.go

@@ -0,0 +1,121 @@
+package main
+
+import (
+	"crypto/tls"
+	"crypto/x509"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// novox/hq 04-ISSUES/146. The bootstrap could not make the bus a certificate: it asked an image for
+// `openssl` and the image it asks has none. What replaces it is this command, so what is checked is
+// what the bootstrap needs from it — a pair a TLS server can actually load, made once and only once.
+
+func TestTheBusCertificateLoadsAsAServersWould(t *testing.T) {
+	into := t.TempDir()
+	if err := busCertificate([]string{"--into", into}); err != nil {
+		t.Fatalf("the bus could not be given a certificate: %v", err)
+	}
+
+	// The check a cheaper test would not make. The key was present and valid and the server could
+	// not start, once, because nothing loaded the pair the way a server loads it
+	// (novox/hq 04-ISSUES/014).
+	pair, err := tls.LoadX509KeyPair(filepath.Join(into, "tls.crt"), filepath.Join(into, "tls.key"))
+	if err != nil {
+		t.Fatalf("a TLS server cannot load what was written: %v", err)
+	}
+	leaf := pair.Leaf
+	if leaf == nil {
+		if leaf, err = x509.ParseCertificate(pair.Certificate[0]); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := leaf.VerifyHostname("mesh-broker"); err != nil {
+		t.Errorf("the certificate is not for the name the bus is reached by: %v", err)
+	}
+	if len(leaf.IPAddresses) == 0 || leaf.IPAddresses[0].String() != "127.0.0.1" {
+		t.Errorf("the certificate does not cover the loopback address the foundation dials: %v", leaf.IPAddresses)
+	}
+
+	// The key is not readable by anything else on the machine; the certificate is public and is.
+	key, err := os.Stat(filepath.Join(into, "tls.key"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if key.Mode().Perm() != 0o600 {
+		t.Errorf("the key is %v", key.Mode().Perm())
+	}
+	crt, err := os.Stat(filepath.Join(into, "tls.crt"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if crt.Mode().Perm() != 0o644 {
+		t.Errorf("the certificate is %v, which the server runs as another user cannot read", crt.Mode().Perm())
+	}
+}
+
+// **Made once.** The step is applied again on every reconcile, and a second certificate is one the
+// hosts that pinned the first no longer believe (novox/hq ADR 0004).
+func TestTheBusCertificateIsMadeOnce(t *testing.T) {
+	into := t.TempDir()
+	if err := busCertificate([]string{"--into", into}); err != nil {
+		t.Fatal(err)
+	}
+	first, err := os.ReadFile(filepath.Join(into, "tls.crt"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := busCertificate([]string{"--into", into}); err != nil {
+		t.Fatal(err)
+	}
+	again, err := os.ReadFile(filepath.Join(into, "tls.crt"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(first) != string(again) {
+		t.Fatal("running it twice replaced the certificate every host had pinned")
+	}
+}
+
+// The verify half: false before, true after, which is what makes the bootstrap run the step at all.
+func TestTheCheckIsFalseUntilThereIsAPair(t *testing.T) {
+	into := t.TempDir()
+	if err := busCertificate([]string{"--check", "--into", into}); err == nil {
+		t.Fatal("an empty directory reported a usable certificate")
+	}
+	if err := busCertificate([]string{"--into", into}); err != nil {
+		t.Fatal(err)
+	}
+	if err := busCertificate([]string{"--check", "--into", into}); err != nil {
+		t.Fatalf("the certificate it just made does not satisfy its own check: %v", err)
+	}
+}
+
+// A half-written pair is not a pair. An interrupted bootstrap leaves exactly this, and a step that
+// called it done would hand the server a certificate with no key and report success.
+func TestACertificateWithoutItsKeyIsNotUsable(t *testing.T) {
+	into := t.TempDir()
+	if err := busCertificate([]string{"--into", into}); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Remove(filepath.Join(into, "tls.key")); err != nil {
+		t.Fatal(err)
+	}
+	if err := busCertificate([]string{"--check", "--into", into}); err == nil {
+		t.Fatal("a certificate with no key passed the check")
+	}
+	if err := busCertificate([]string{"--into", into}); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := tls.LoadX509KeyPair(filepath.Join(into, "tls.crt"), filepath.Join(into, "tls.key")); err != nil {
+		t.Fatalf("it did not replace the unusable pair: %v", err)
+	}
+}
+
+func TestWhereToWriteIsRequired(t *testing.T) {
+	if err := busCertificate(nil); err == nil || !strings.Contains(err.Error(), "--into") {
+		t.Fatalf("it did not ask where to write: %v", err)
+	}
+}

cmd/mesh-controller/calls_verb_test.go

@@ -0,0 +1,81 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// consoleWaits is how long the console waits for an answer (mesh-tools node-tools/internal/bus
+// RequestTimeout) — the shortest wait of a caller the mesh ships.
+const consoleWaits = 30 * time.Second
+
+// **A call's one answer is never later than its caller or the bus allow** (novox/hq issue 265): a
+// holder answers within AnswerWithin, which must be inside both the console's wait and the window the
+// bus gives an answer. Before, the console waited 30s, the bus 60s, and a push ran as long as it ran.
+func TestAVerbAnswersInsideEveryWaitOnIt(t *testing.T) {
+	if link.AnswerWithin >= consoleWaits/2 {
+		t.Errorf("a call answers within %s: not well inside the console's %s", link.AnswerWithin, consoleWaits)
+	}
+	if link.AnswerWithin >= broker.ResponseTTL {
+		t.Errorf("a call answers within %s, after the bus stops permitting an answer at %s", link.AnswerWithin, broker.ResponseTTL)
+	}
+}
+
+// A push — named or through command — answers before it runs: it sends the machine holding the bus
+// first, and the broker reloading its user list forgets the answer it was about to permit.
+func TestAPushAnswersBeforeItSends(t *testing.T) {
+	for _, c := range []struct {
+		verb string
+		args map[string]any
+		want bool
+	}{
+		{"push", map[string]any{"node": "anchor"}, true},
+		{"push", map[string]any{}, true},
+		{"command", map[string]any{"command": "push anchor"}, true},
+		{"command", map[string]any{"command": "push --behind"}, true},
+		{"command", map[string]any{"command": "builds"}, false},
+		{"status", map[string]any{}, false},
+		{"assign", map[string]any{"node": "anchor", "module": "m"}, false},
+	} {
+		argv, err := argvFor(c.verb, c.args)
+		if err != nil {
+			t.Fatalf("%s %v: %v", c.verb, c.args, err)
+		}
+		if got := answersFirst(argv); got != c.want {
+			t.Errorf("%s %v answers first: %v, want %v", c.verb, c.args, got, c.want)
+		}
+	}
+}
+
+// `calls` is served, takes a call's id and nothing else, and says plainly when it holds no such call.
+func TestCallsIsServedAndSaysWhatItKeeps(t *testing.T) {
+	handlers, behind, err := seatToolHandlers()
+	if err != nil || len(behind) != 0 {
+		t.Fatalf("%v %v", behind, err)
+	}
+	calls, ok := handlers["calls"]
+	if !ok {
+		t.Fatal("calls is not served")
+	}
+	if _, err := calls(context.Background(), json.RawMessage(`{"node":"anchor"}`)); err == nil ||
+		!strings.Contains(err.Error(), `"node"`) {
+		t.Errorf("calls took an argument it does not declare: %v", err)
+	}
+	if _, err := calls(context.Background(), json.RawMessage(`{"call":"call-0-0"}`)); err == nil ||
+		!strings.Contains(err.Error(), "not across a restart") {
+		t.Errorf("an unknown call was not said plainly: %v", err)
+	}
+	got, err := calls(context.Background(), json.RawMessage(`{}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, listed := got.(map[string]any)["calls"]; !listed {
+		t.Errorf("calls answered %v", got)
+	}
+}

cmd/mesh-controller/check.go

@@ -0,0 +1,147 @@
+package main
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// moduleCheck judges manifests where they are written, with no mesh (novox/hq ADR 0037, issue 148).
+//
+// **The same functions registration runs, and nothing the command line adds** (ADR 0035): the strict
+// parse with every per-manifest problem, then the rules no single manifest can be judged against,
+// over exactly the manifests given. Somebody describing their own application in their own
+// repository runs this before pushing and finds out there, rather than when a running mesh refuses
+// the registration or, later, when a machine applies something that resolved and should not have.
+//
+// **What it cannot know without a store, it says.** The mesh's own seat set is the store's (ADR
+// 0122); this binary carries a compiled copy that the store overrides when loaded, so a claim on a
+// mesh seat is judged fully only at registration. A seat another module declares is unknown unless
+// that module's manifest is passed too. Both are printed as a note, not as a problem — a check that
+// refused what it could not see would teach people to ignore it.
+func moduleCheck(paths []string, out io.Writer) error {
+	if len(paths) == 0 {
+		return errors.New("module check <manifest.json>... — one file per module; pass every " +
+			"manifest of a repository together so the rules between them are checked too")
+	}
+	shelf := catalogue.Shelf{}
+	faulted := map[string]bool{}
+	failed := 0
+	for _, path := range paths {
+		raw, err := os.ReadFile(path)
+		if err != nil {
+			fmt.Fprintf(out, "%s: %v\n", path, err)
+			failed++
+			continue
+		}
+		m, err := catalogue.ParseManifest(raw)
+		if err != nil {
+			fmt.Fprintf(out, "%s: %v\n", path, err)
+			failed++
+			continue
+		}
+		if first, twice := shelf[m.Module]; twice {
+			_ = first
+			fmt.Fprintf(out, "%s: %s was already given; two manifests name one module\n", path, m.Module)
+			failed++
+			continue
+		}
+		// A definition names no installation (novox/hq ADR 0112, ADR 0155): judged here, in the
+		// catalogue-wide test, and at registration, which refuses in the same words.
+		if named := catalogue.InstallationProblems(m); len(named) > 0 {
+			for _, p := range named {
+				fmt.Fprintf(out, "%s: %s\n", path, p)
+			}
+			failed += len(named)
+			faulted[m.Module] = true
+		}
+		shelf[m.Module] = m
+	}
+
+	// Between the manifests: a seat declared twice, a use of a seat nothing declares, a claim on
+	// a seat that does not exist. Run only over what parsed, because a problem inside one manifest
+	// has already been said and would be said again here in a worse form.
+	problems := catalogue.CatalogueProblems(shelf)
+	sort.Strings(problems)
+	for _, p := range problems {
+		fmt.Fprintln(out, p)
+	}
+	failed += len(problems)
+
+	var names []string
+	for name := range shelf {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		m := shelf[name]
+		if faulted[name] {
+			continue
+		}
+		fmt.Fprintf(out, "%s: ok", name)
+		if n := len(m.Tools); n > 0 {
+			fmt.Fprintf(out, ", %d tool(s)", n)
+		}
+		if len(m.Invokes) > 0 {
+			fmt.Fprintf(out, ", invokes %s", joinInvokes(m.Invokes))
+		}
+		// The state it keeps and reads (novox/hq ADR 0201), so a reviewer sees what lands on the bus.
+		if len(m.State) > 0 {
+			kept := make([]string, 0, len(m.State))
+			for _, s := range m.State {
+				kept = append(kept, s.Name)
+			}
+			fmt.Fprintf(out, ", keeps state %s", strings.Join(kept, ", "))
+		}
+		if len(m.Reads) > 0 {
+			fmt.Fprintf(out, ", reads %s", strings.Join(m.Reads, ", "))
+		}
+		fmt.Fprintln(out)
+	}
+	if failed > 0 {
+		return fmt.Errorf("%d problem(s) in %d manifest(s)", failed, len(paths))
+	}
+	fmt.Fprintf(out, "%d manifest(s) checked. Judged against the seats this binary carries; a claim on "+
+		"one of the mesh's own seats is judged fully at registration, and a seat declared by a "+
+		"module not given here reads as unknown\n", len(paths))
+	return nil
+}
+
+func joinInvokes(invokes []string) string {
+	if len(invokes) == 1 && invokes[0] == "*" {
+		return "every tool"
+	}
+	s := ""
+	for i, t := range invokes {
+		if i > 0 {
+			s += ", "
+		}
+		s += t
+	}
+	return s
+}
+
+// manifestsUnder lists every module.json below a directory, for `module check <dir>`.
+func manifestsUnder(dir string) ([]string, error) {
+	var found []string
+	err := filepath.WalkDir(dir, func(path string, d os.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if d.IsDir() && (d.Name() == "node_modules" || d.Name() == ".git" || d.Name() == "dist") {
+			return filepath.SkipDir
+		}
+		if !d.IsDir() && d.Name() == "module.json" {
+			found = append(found, path)
+		}
+		return nil
+	})
+	sort.Strings(found)
+	return found, err
+}

cmd/mesh-controller/check_test.go

@@ -0,0 +1,94 @@
+package main
+
+import (
+	"bytes"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"testing"
+)
+
+// The check anybody can run is the check registration runs (novox/hq issue 148, ADR 0037): a manifest
+// with a known fault is named, and one without passes, with no store opened.
+func TestModuleCheckNamesAFaultAndNeedsNoMesh(t *testing.T) {
+	dir := t.TempDir()
+	good := filepath.Join(dir, "good.json")
+	bad := filepath.Join(dir, "bad.json")
+	os.WriteFile(good, []byte(`{"module":"shop","version":"1","tools":["price"],"invokes":["mesh-catalog.catalog_modules"]}`), 0o600)
+	os.WriteFile(bad, []byte(`{"module":"till","version":"1","invokes":["shop"]}`), 0o600)
+
+	var out bytes.Buffer
+	if err := moduleCheck([]string{good}, &out); err != nil {
+		t.Fatalf("a sound manifest was refused: %v\n%s", err, out.String())
+	}
+	if !strings.Contains(out.String(), "shop: ok, 1 tool(s), invokes mesh-catalog.catalog_modules") {
+		t.Fatalf("the report does not say what it checked:\n%s", out.String())
+	}
+
+	out.Reset()
+	err := moduleCheck([]string{good, bad}, &out)
+	if err == nil {
+		t.Fatal("a manifest invoking a module and no tool passed")
+	}
+	if !strings.Contains(out.String(), `till invokes "shop", which does not name a tool`) {
+		t.Fatalf("the fault is not named in the manifest's words:\n%s", out.String())
+	}
+}
+
+// The rules between manifests run over what was given together: a seat two modules declare is
+// refused, which no single-manifest check can see.
+func TestModuleCheckJudgesBetweenTheManifestsGiven(t *testing.T) {
+	dir := t.TempDir()
+	a := filepath.Join(dir, "a.json")
+	b := filepath.Join(dir, "b.json")
+	os.WriteFile(a, []byte(`{"module":"a","version":"1","seats":[{"name":"printer","scope":"mesh"}]}`), 0o600)
+	os.WriteFile(b, []byte(`{"module":"b","version":"1","seats":[{"name":"printer","scope":"mesh"}]}`), 0o600)
+	var out bytes.Buffer
+	if err := moduleCheck([]string{a, b}, &out); err == nil {
+		t.Fatalf("two declarations of one seat passed:\n%s", out.String())
+	}
+	if !strings.Contains(out.String(), "a seat name means one protocol") {
+		t.Fatalf("the cross-manifest rule was not the one named:\n%s", out.String())
+	}
+}
+
+// The real catalogue passes the command, the way it passes the test that used to be the only check.
+func TestModuleCheckPassesTheCatalogue(t *testing.T) {
+	root := filepath.Join("..", "..", "..", "mesh-catalog", "modules")
+	if _, err := os.Stat(root); err != nil {
+		t.Skipf("catalogue sibling not present: %v", err)
+	}
+	paths, err := manifestsUnder(root)
+	if err != nil || len(paths) == 0 {
+		t.Fatalf("no manifests under %s: %v", root, err)
+	}
+	var out bytes.Buffer
+	if err := moduleCheck(paths, &out); err != nil {
+		t.Fatalf("the catalogue does not pass its own check: %v\n%s", err, out.String())
+	}
+}
+
+func TestRegistrationRefusesADefinitionNamingAnInstallation(t *testing.T) {
+	// novox/hq ADR 0155: the check moves to registration once the catalogue passes it. Both
+	// ways in — `module add` and a build's result — go through this, and a name declared on
+	// purpose passes with its reason.
+	named := catalogue.Manifest{Module: "idp", Resources: []map[string]any{
+		{"id": "server", "type": "container", "image": "x@sha256:aa",
+			"env": map[string]any{"KC_HOSTNAME": "https://login.mesh-one.be"}},
+	}}
+	err := namesNoInstallation(named)
+	if err == nil || !strings.Contains(err.Error(), "login.mesh-one.be") ||
+		!strings.Contains(err.Error(), catalogue.NamesOnPurpose) {
+		t.Fatalf("a definition naming an installation is refused with the name and the way out; got %v", err)
+	}
+	meant := catalogue.Manifest{Module: "site", Resources: []map[string]any{
+		{"id": "server", "type": "container", "image": "registry.mesh-one.be/org/site@sha256:cc",
+			catalogue.NamesOnPurpose: map[string]any{
+				"registry.mesh-one.be": "built outside the mesh until its repository is a build source here"}},
+	}}
+	if err := namesNoInstallation(meant); err != nil {
+		t.Fatalf("a name declared on purpose passes; got %v", err)
+	}
+}

cmd/mesh-controller/collect.go

@@ -0,0 +1,175 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/artifacts"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// Letting the artifact store go of what the mesh no longer keeps (novox/hq ADR 0189, issue 108).
+//
+// **Run where the records change.** A build is the moment new bytes landed in the store and the
+// moment the keep set moved, so it is the moment to say what may go — and it needs no timer of
+// its own. Reclaiming the bytes is the store's own nightly step; this only decides.
+//
+// Never fatal to a build. The build succeeded, the module is registered, and a store that could
+// not be reached is a thing to say rather than a reason to undo any of that. The next build asks
+// again, and the references it could not collect are still uncollected, so nothing is lost by
+// having failed.
+
+// collect asks the store to let go of everything the mesh made and no longer keeps, and records
+// what it let go of. Says what it did and what it could not; returns nothing, because nothing
+// upstream should branch on it.
+func collect(ctx context.Context, inv *inventory.Inventory) {
+	references, err := inv.ToCollect(ctx)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not work out what the artifact store may let go of: %v\n", err)
+		return
+	}
+	kept, err := inv.KeptArchives(ctx)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not work out which archives the artifact store keeps: %v\n", err)
+		return
+	}
+	if len(references) == 0 && len(kept) == 0 {
+		return
+	}
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not read the catalogue to find the artifact store: %v\n", err)
+		return
+	}
+	// As the mesh reaches it from the network. Empty means the store is not on the network — on a
+	// mesh being raised it is not yet, and there the store holds one build of anything and has
+	// nothing to collect.
+	address, err := artifactStoreAddress(ctx, inv, shelf, "")
+	if err != nil || address == "" {
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "could not find the artifact store to collect from: %v\n", err)
+		}
+		return
+	}
+
+	// **Bounded, because this runs inside somebody's build.** The first sweep of a mesh that has
+	// never collected has the whole history to get through, and a person waiting on `build` should
+	// not pay for it. Two bounds, and what is left over is simply offered again next time —
+	// builds are frequent, and the point is that the store stops growing, not that it empties
+	// tonight.
+	within, stop := context.WithTimeout(ctx, sweepBudget)
+	defer stop()
+	store := artifacts.Store{Address: address}
+
+	// **Hold before letting go** (novox/hq issue 253). The store's collector keeps only what a
+	// manifest names, and archives were published as bare blobs, so every kept archive is first
+	// held by its manifest — which backfills the ones published before holders, a few at a time
+	// as builds come, and is two HEADs each once done. A kept archive that could not be held stops
+	// the sweep before it deletes anything: "everything kept is held" is the precondition the
+	// collector's safety rests on, and a store refusing a hold would refuse the deletes too.
+	wrote, missing, err := holdKept(within, store, kept)
+	if wrote > 0 {
+		fmt.Fprintf(os.Stderr, "the artifact store now holds %d more kept archive(s) by a manifest\n", wrote)
+	}
+	if missing > 0 {
+		fmt.Fprintf(os.Stderr, "%d archive(s) the mesh keeps are not in the artifact store at all; "+
+			"`collection` lists them\n", missing)
+	}
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "not every kept archive could be held, so nothing was let go: %v\n", err)
+		return
+	}
+
+	var done []string
+	var left, skipped int
+	for i, reference := range references {
+		if i >= mostPerSweep || within.Err() != nil {
+			left = len(references) - i
+			break
+		}
+		err := store.LetGo(within, reference)
+		if err == nil || errors.Is(err, artifacts.Gone) {
+			// Gone is the outcome wanted, already true. Recorded so the next sweep does not ask
+			// again for ever.
+			done = append(done, reference)
+			continue
+		}
+		if errors.Is(err, artifacts.ErrNotOurs) {
+			// **A fact about this record, so this record is skipped** (novox/hq issue 226). Not
+			// marked collected — the mesh did not remove it and should not claim to — and not a
+			// reason to stop, because the store was never asked. One of these at the front of
+			// the oldest-first order ended every sweep until this.
+			skipped++
+			if skipped == 1 {
+				fmt.Fprintf(os.Stderr,
+					"the sweep will not address %s and went on: %v\n", reference, err)
+			}
+			continue
+		}
+		// **Stopped at the first refusal by the STORE, not pushed through.** A store that refuses
+		// one refuses all of them — deletion disabled, the store down, the network gone — so
+		// going on would be a hundred identical failures and a hundred identical log lines in
+		// front of whoever was building something.
+		fmt.Fprintf(os.Stderr, "the artifact store kept %s, so nothing more was asked of it: %v\n",
+			reference, err)
+		left = len(references) - i
+		break
+	}
+
+	if len(done) > 0 {
+		// Recorded outside `within`: the deletions happened, and losing the record of them because
+		// the sweep ran out of budget would mean asking about them again for ever.
+		if err := inv.MarkCollected(ctx, done); err != nil {
+			fmt.Fprintf(os.Stderr, "the store let go of %d artifact(s) and the record of it did not keep: %v\n",
+				len(done), err)
+			return
+		}
+		fmt.Fprintf(os.Stderr, "the artifact store let go of %d artifact(s) the mesh no longer keeps\n",
+			len(done))
+	}
+	if left > 0 {
+		fmt.Fprintf(os.Stderr, "%d more to collect; the next build asks again\n", left)
+	}
+	if skipped > 0 {
+		fmt.Fprintf(os.Stderr, "%d artifact(s) the sweep will not address were skipped\n", skipped)
+	}
+}
+
+// holdKept holds every kept archive by its manifest, stopping at the first refusal by the store.
+// Answers how many holders it wrote and how many kept archives the store does not have.
+//
+// A missing archive is counted rather than fatal: there is nothing to hold, and that is a fact
+// for an operator to read (`collection`), not a reason to stop collecting what is not kept. A
+// reference the store cannot be asked about is skipped as the deletion loop skips one
+// (novox/hq issue 226).
+func holdKept(ctx context.Context, store artifacts.Store, kept []string) (wrote, missing int, err error) {
+	for _, reference := range kept {
+		if err := ctx.Err(); err != nil {
+			return wrote, missing, fmt.Errorf("ran out of time before %s: %w", reference, err)
+		}
+		did, err := store.Hold(ctx, reference)
+		switch {
+		case err == nil:
+			if did {
+				wrote++
+			}
+		case errors.Is(err, artifacts.Gone):
+			missing++
+		case errors.Is(err, artifacts.ErrNotOurs):
+		default:
+			return wrote, missing, fmt.Errorf("holding %s: %w", reference, err)
+		}
+	}
+	return wrote, missing, nil
+}
+
+// mostPerSweep is how many artifacts one sweep will ask about. Enough that a mesh building
+// several times a day converges within days of this landing; small enough that no single build
+// waits on the whole backlog.
+const mostPerSweep = 200
+
+// sweepBudget is the longest a sweep will keep a build waiting.
+const sweepBudget = 60 * time.Second

cmd/mesh-controller/collection.go

@@ -0,0 +1,166 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"os"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/artifacts"
+)
+
+// What the artifact store keeps, whether each kept archive is held, and what the sweep may let go
+// (novox/hq issue 253, ADR 0189).
+//
+// **The question to answer before the store's collector runs for real.** The collector deletes
+// every blob no manifest names, and archives were published as bare blobs, so the nightly step
+// runs `--dry-run` until every archive the mesh keeps is held by its manifest. The sweep holds
+// them as builds come; this says how far that has got — "0 unheld" is the number that lets the
+// dry run go.
+//
+// Reads and changes nothing: each kept archive is asked about with HEADs only. Reached over the
+// console through the mesh-controller seat's `command` verb (`collection --json`), which needs no
+// new verb in the seat's row.
+
+type collectionReport struct {
+	// Store is the artifact store as this machine reached it; empty when it is not on the network.
+	Store string `json:"store"`
+	// KeptArchives is how many archives the mesh keeps, for either reason.
+	KeptArchives int `json:"kept_archives"`
+	// Held is how many of them the store holds by their manifest.
+	Held int `json:"held"`
+	// Unheld are the kept archives the collector would delete tonight if it ran for real.
+	Unheld []string `json:"unheld"`
+	// Missing are kept archives the store does not have at all.
+	Missing []string `json:"missing"`
+	// Unasked is how many could not be asked about, and why the asking stopped.
+	Unasked int    `json:"unasked"`
+	Stopped string `json:"stopped,omitempty"`
+	// Eligible is what the sweep may let go of: made by the mesh, kept for no reason, not yet
+	// collected — split by kind.
+	Eligible         int `json:"eligible"`
+	EligibleImages   int `json:"eligible_images"`
+	EligibleArchives int `json:"eligible_archives"`
+	// SafeToCollect is whether every kept archive was asked about and every one is held.
+	SafeToCollect bool `json:"safe_to_collect"`
+}
+
+func collectionCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("collection", flag.ContinueOnError)
+	asJSON := set.Bool("json", false, "answer as JSON")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 0 {
+		return errors.New("collection [--json]")
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	kept, err := inv.KeptArchives(ctx)
+	if err != nil {
+		return err
+	}
+	eligible, err := inv.ToCollect(ctx)
+	if err != nil {
+		return err
+	}
+	report := collectionReport{KeptArchives: len(kept), Eligible: len(eligible), Unheld: []string{}, Missing: []string{}}
+	for _, reference := range eligible {
+		if strings.Contains(reference, "/blobs/") {
+			report.EligibleArchives++
+		} else {
+			report.EligibleImages++
+		}
+	}
+
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return err
+	}
+	report.Store, err = artifactStoreAddress(ctx, inv, shelf, "")
+	if err != nil {
+		return err
+	}
+	if report.Store == "" {
+		report.Unasked = len(kept)
+		report.Stopped = "this mesh has no artifact store on its network"
+	} else {
+		report.Unasked, report.Stopped = askHeld(ctx, artifacts.Store{Address: report.Store}, kept, &report)
+	}
+	report.SafeToCollect = report.Unasked == 0 && len(report.Unheld) == 0
+
+	if *asJSON {
+		encoder := json.NewEncoder(os.Stdout)
+		encoder.SetIndent("", "  ")
+		return encoder.Encode(report)
+	}
+	printCollection(report)
+	return nil
+}
+
+// askHeld asks the store about each kept archive, stopping at the first answer that is not about
+// the archive: a store that cannot be reached for one cannot be for the next, and a page of
+// identical failures says less than one line.
+func askHeld(ctx context.Context, store artifacts.Store, kept []string, report *collectionReport) (int, string) {
+	for i, reference := range kept {
+		held, err := store.Held(ctx, reference)
+		switch {
+		case err == nil && held:
+			report.Held++
+		case err == nil:
+			report.Unheld = append(report.Unheld, reference)
+		case errors.Is(err, artifacts.Gone):
+			report.Missing = append(report.Missing, reference)
+		case errors.Is(err, artifacts.ErrNotOurs):
+			// KeptArchives names only the mesh's own; counted as unasked if one ever is not.
+			report.Unasked++
+		default:
+			return report.Unasked + len(kept) - i, fmt.Sprintf("asking about %s: %v", reference, err)
+		}
+	}
+	return report.Unasked, ""
+}
+
+func printCollection(r collectionReport) {
+	store := r.Store
+	if store == "" {
+		store = "(not on the network)"
+	}
+	fmt.Printf("artifact store       %s\n", store)
+	fmt.Printf("kept archives        %d\n", r.KeptArchives)
+	fmt.Printf("  held               %d\n", r.Held)
+	fmt.Printf("  unheld             %d\n", len(r.Unheld))
+	fmt.Printf("  missing            %d\n", len(r.Missing))
+	if r.Unasked > 0 {
+		fmt.Printf("  not asked          %d (%s)\n", r.Unasked, r.Stopped)
+	}
+	fmt.Printf("eligible to let go   %d (%d images, %d archives)\n", r.Eligible, r.EligibleImages, r.EligibleArchives)
+	if len(r.Unheld) > 0 {
+		fmt.Println("\nunheld — the store's collector would delete these; the next build's sweep holds them:")
+		for _, reference := range r.Unheld {
+			fmt.Printf("  %s\n", reference)
+		}
+	}
+	if len(r.Missing) > 0 {
+		fmt.Println("\nmissing — kept by the mesh, not in the store:")
+		for _, reference := range r.Missing {
+			fmt.Printf("  %s\n", reference)
+		}
+	}
+	fmt.Println()
+	if r.SafeToCollect {
+		fmt.Println("every kept archive is held: the store's collector may run for real")
+	} else {
+		fmt.Println("NOT every kept archive is known to be held: keep the store's collector on --dry-run")
+	}
+}

cmd/mesh-controller/credential_on_assign_test.go

@@ -0,0 +1,90 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// A fresh assignment is pushed before its credential exists (novox/hq issue 203): `assign` recorded
+// the module, `push` sealed a random own secret where the bus credential belongs, and the process
+// crash-looped until a person ran `module issue` and pushed again. Now assigning a module that speaks
+// on the bus issues its credential in the same act — or, when the bus cannot be reached from here,
+// says which verb to run — and a push never seals a placeholder in a credential's place.
+
+func aTalker() catalogue.Manifest {
+	return catalogue.Manifest{Module: "talker", Version: "1",
+		OwnSecrets: catalogue.OwnSecrets{"broker": {Path: "/var/lib/mesh/talker/broker"}},
+		Resources: []map[string]any{
+			{"id": "state", "type": "directory", "path": "/var/lib/mesh/talker", "mode": "0700"},
+		}}
+}
+
+func TestAssigningAModuleThatSpeaksOnTheBusNamesItsCredential(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, aTalker())
+
+	// No bus is known to this process, so the credential cannot be issued here: the assignment
+	// stands and says exactly what must happen before a push — never silently.
+	said, err := assign(ctx, open, "laptop", "talker")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(said, "module issue talker --node laptop") {
+		t.Fatalf("an assignment whose credential could not be issued does not name the verb:\n%s", said)
+	}
+
+	// And the push refuses to send it, naming the same verb, rather than sealing a placeholder.
+	plan, settings, err := planFor(ctx, open, "laptop")
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = declarationFor(ctx, open, "laptop", plan, settings)
+	if err == nil {
+		t.Fatal("a push sealed a placeholder where talker's bus credential belongs")
+	}
+	if !strings.Contains(err.Error(), "module issue talker --node laptop") || !strings.Contains(err.Error(), "issue 203") {
+		t.Fatalf("the refusal does not say what to run: %v", err)
+	}
+
+	// Once the user is minted, the push goes on to the credential the mesh sealed, and re-assigning
+	// does not mint again: a credential rotates on purpose, never by habit.
+	if _, err := open.inventory.MintBusPassword(ctx, inventory.BusUser{
+		Username: "laptop.talker", Kind: inventory.BusModule, Node: "laptop", Module: "talker"}); err != nil {
+		t.Fatal(err)
+	}
+	hash, _, err := open.inventory.BusUserHash(ctx, "laptop.talker")
+	if err != nil {
+		t.Fatal(err)
+	}
+	said, err = assign(ctx, open, "laptop", "talker")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(said, "module issue") {
+		t.Fatalf("a module with a minted credential was told to issue one:\n%s", said)
+	}
+	again, _, err := open.inventory.BusUserHash(ctx, "laptop.talker")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if again != hash {
+		t.Fatal("re-assigning rotated the credential")
+	}
+}
+
+// A module that declares no broker secret is left alone: nothing to issue, nothing said.
+func TestAssigningAModuleThatDoesNotSpeakSaysNothingOfCredentials(t *testing.T) {
+	open := aMesh(t)
+	register(t, open, helloWeb())
+	said, err := assign(t.Context(), open, "laptop", "hello-web")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(said, "credential") {
+		t.Fatalf("a module without a broker secret was told about credentials:\n%s", said)
+	}
+}

cmd/mesh-controller/delivery_order_test.go

@@ -0,0 +1,145 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"reflect"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// recordingDelivery is a delivery that writes down what was done, in order, and fails where told.
+type recordingDelivery struct {
+	did        []string
+	grantErr   error
+	declareErr error
+}
+
+func (r *recordingDelivery) grant(_ context.Context, sending []readyNode) error {
+	for _, s := range sending {
+		r.did = append(r.did, "grant "+s.node)
+	}
+	return r.grantErr
+}
+
+func (r *recordingDelivery) declare(_ context.Context, s readyNode, _ []byte) (string, error) {
+	if r.declareErr != nil {
+		return "", r.declareErr
+	}
+	r.did = append(r.did, "declare "+s.node)
+	return "digest-" + s.node, nil
+}
+
+func ready(names ...string) []readyNode {
+	var out []readyNode
+	for _, n := range names {
+		out = append(out, readyNode{node: n, declared: sendable{Resources: []map[string]any{{"id": "x"}}}})
+	}
+	return out
+}
+
+// novox/hq issue 249: the grants that come with a module's new declarations are issued before any
+// machine is sent the code that uses them — except the machine holding the bus, whose declaration
+// carries the controller's own right to issue them, and goes first.
+func TestGrantsAreIssuedBeforeTheDeclarations(t *testing.T) {
+	d := &recordingDelivery{}
+	digests, err := deliver(t.Context(), d, "", ready("anchor", "laptop"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := []string{"grant anchor", "grant laptop", "declare anchor", "declare laptop"}
+	if !reflect.DeepEqual(d.did, want) {
+		t.Fatalf("delivered in the order %v, wanted %v", d.did, want)
+	}
+	if digests["anchor"] != "digest-anchor" || digests["laptop"] != "digest-laptop" {
+		t.Fatalf("the digests sent were not answered: %v", digests)
+	}
+
+	held := &recordingDelivery{}
+	if _, err := deliver(t.Context(), held, "broker", ready("broker", "anchor")); err != nil {
+		t.Fatal(err)
+	}
+	want = []string{"declare broker", "grant broker", "grant anchor", "declare anchor"}
+	if !reflect.DeepEqual(held.did, want) {
+		t.Fatalf("with the bus's machine in the send: %v, wanted %v", held.did, want)
+	}
+}
+
+// A grant that cannot be issued holds back the machines it concerns and is an error the caller
+// retries on — never "until the next push" — and the bus's own machine is sent regardless, so the
+// grant that would let the controller issue memberships is never held behind them.
+func TestAGrantThatFailsHoldsBackWhatItConcerns(t *testing.T) {
+	// A failure naming no machine (the buckets): everything but the bus's machine.
+	d := &recordingDelivery{grantErr: errors.New("the bus refused the bucket")}
+	_, err := deliver(t.Context(), d, "broker", ready("broker", "anchor", "laptop"))
+	if err == nil || !errors.Is(err, errGrants) || !strings.Contains(err.Error(), "the bus refused the bucket") ||
+		!strings.Contains(err.Error(), "anchor, laptop not sent") {
+		t.Fatalf("a failed grant was not said as the send's failure: %v", err)
+	}
+	if want := []string{"declare broker", "grant broker", "grant anchor", "grant laptop"}; !reflect.DeepEqual(d.did, want) {
+		t.Fatalf("delivered %v, wanted the bus's machine alone", d.did)
+	}
+
+	// A membership that failed for one machine: that machine alone.
+	one := &recordingDelivery{grantErr: &grantsRefused{nodes: map[string]error{"laptop": errors.New("no")}}}
+	_, err = deliver(t.Context(), one, "", ready("anchor", "laptop"))
+	if !errors.Is(err, errGrants) || !strings.Contains(err.Error(), "laptop not sent") {
+		t.Fatalf("one machine's refused membership was not said: %v", err)
+	}
+	if want := []string{"grant anchor", "grant laptop", "declare anchor"}; !reflect.DeepEqual(one.did, want) {
+		t.Fatalf("delivered %v, wanted anchor sent and laptop held back", one.did)
+	}
+
+	// The announced upgrade that hit it is asked again.
+	if !errors.Is(askAgainOnGrants(err), link.ErrTryAgain) {
+		t.Fatal("an announcement whose send stopped at its grants is not asked again")
+	}
+	if other := errors.New("laptop could not be resolved"); errors.Is(askAgainOnGrants(other), link.ErrTryAgain) {
+		t.Fatal("any failure is asked again, not only a grant's")
+	}
+
+	// Nothing to send is nothing granted either.
+	none := &recordingDelivery{grantErr: errors.New("never asked")}
+	if _, err := deliver(t.Context(), none, "", nil); err != nil || len(none.did) != 0 {
+		t.Fatalf("an empty send granted or failed: %v %v", none.did, err)
+	}
+}
+
+// Whether the bus's machine goes first is read from the user list alone, by its digest.
+func TestTheBusMachineIsBehindByItsUserListAlone(t *testing.T) {
+	list := "users: [a, b]"
+	if userListBehind(list, digestOf([]byte(list))) {
+		t.Fatal("the list it was sent reads as behind")
+	}
+	if !userListBehind(list, digestOf([]byte("users: [a]"))) || !userListBehind(list, "") {
+		t.Fatal("a changed or never-sent list reads as current")
+	}
+	if userListBehind("", "") {
+		t.Fatal("a machine sent no list reads as behind")
+	}
+}
+
+// The machine holding the bus goes first: its declaration carries the user list the new grants are
+// checked against. Among the machines it is moved to the front; not among them it is added only
+// when it is behind.
+func TestTheMachineHoldingTheBusIsSentFirst(t *testing.T) {
+	for _, c := range []struct {
+		what   string
+		names  []string
+		holder string
+		behind bool
+		want   []string
+	}{
+		{"among them", []string{"ace", "g14", "novox"}, "novox", false, []string{"novox", "ace", "g14"}},
+		{"not among them, behind", []string{"ace", "g14"}, "novox", true, []string{"novox", "ace", "g14"}},
+		{"not among them, current", []string{"ace", "g14"}, "novox", false, []string{"ace", "g14"}},
+		{"nothing holds the bus", []string{"ace", "g14"}, "", true, []string{"ace", "g14"}},
+		{"only it", []string{"novox"}, "novox", false, []string{"novox"}},
+	} {
+		if got := brokerFirst(c.names, c.holder, c.behind); !reflect.DeepEqual(got, c.want) {
+			t.Errorf("%s: sent in the order %v, wanted %v", c.what, got, c.want)
+		}
+	}
+}

cmd/mesh-controller/dependents_test.go

@@ -0,0 +1,52 @@
+package main
+
+import (
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// A merge that rebuilds a base rebuilds what stands on it, through every layer, and nothing else
+// (novox/hq issue 186): the runtime image moving means every module built on it moves too, and a
+// module built on one of those moves as well.
+func TestAMergeOfABaseTakesWhatStandsOnItAlong(t *testing.T) {
+	entry := func(name string) inventory.Entry {
+		return inventory.Entry{Manifest: catalogue.Manifest{Module: name}}
+	}
+	entries := []inventory.Entry{entry("mesh-tools"), entry("shop"), entry("shop-plugin"), entry("postgres"), entry("unrelated")}
+	against := map[string][]string{
+		"shop":        {catalogue.ArtifactStoreScheme + "mesh-tools/runtime@sha256:a"},
+		"shop-plugin": {catalogue.ArtifactStoreScheme + "shop/runtime@sha256:b"},
+		"postgres":    {catalogue.ArtifactStoreScheme + "mesh-tools/runtime@sha256:a"},
+		"unrelated":   {catalogue.ArtifactStoreScheme + "alpine/base@sha256:c"},
+	}
+	got := dependentsOf([]inventory.Entry{entry("mesh-tools")}, entries, against)
+	var names []string
+	for _, e := range got {
+		names = append(names, e.Manifest.Module)
+	}
+	want := map[string]bool{"shop": true, "shop-plugin": true, "postgres": true}
+	if len(names) != len(want) {
+		t.Fatalf("rebuilt %v; wanted exactly the three that stand on the runtime, directly or through shop", names)
+	}
+	for _, n := range names {
+		if !want[n] {
+			t.Fatalf("%s was rebuilt and stands on nothing that moved (%v)", n, names)
+		}
+	}
+	// The dependents come in base order when the merge orders them: the runtime, then shop, then
+	// the plugin that stands on shop.
+	ordered := orderByBases(append([]inventory.Entry{entry("mesh-tools")}, got...), against)
+	pos := map[string]int{}
+	for i, e := range ordered {
+		pos[e.Manifest.Module] = i
+	}
+	if !(pos["mesh-tools"] < pos["shop"] && pos["shop"] < pos["shop-plugin"]) {
+		t.Fatalf("not in base order: %v", ordered)
+	}
+	// Nothing moved: nothing follows.
+	if more := dependentsOf(nil, entries, against); len(more) != 0 {
+		t.Fatalf("with nothing moved, %d module(s) were rebuilt", len(more))
+	}
+}

cmd/mesh-controller/dry_run_test.go

@@ -0,0 +1,38 @@
+package main
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A dry run's outcome is looked at, never taken in (novox/hq issue 240). The daemon here holds no
+// store at all, so anything that tried to record or register would fail rather than pass quietly.
+func TestADryRunsOutcomeIsTakenInByNothing(t *testing.T) {
+	err := builds{}.Built(t.Context(), link.BuildResult{
+		ID: "build-1", Repository: "ssh://forge/app.git", Ref: "unreviewed", Module: "app", DryRun: true,
+		Manifest: json.RawMessage(`{"module":"app","version":"1"}`),
+	})
+	if err != nil {
+		t.Fatalf("a dry run's outcome was not simply set aside: %v", err)
+	}
+}
+
+// The mark survives the wire both ways: asked as a dry run, answered as one.
+func TestTheDryRunMarkTravelsWithTheBuild(t *testing.T) {
+	raw, _ := json.Marshal(link.BuildRequest{ID: "build-1", Repository: "r", DryRun: true})
+	var asked link.BuildRequest
+	if err := json.Unmarshal(raw, &asked); err != nil || !asked.DryRun {
+		t.Fatalf("the request lost its dry-run mark: %s", raw)
+	}
+	raw, _ = json.Marshal(link.BuildResult{ID: "build-1", DryRun: true})
+	var answered link.BuildResult
+	if err := json.Unmarshal(raw, &answered); err != nil || !answered.DryRun {
+		t.Fatalf("the outcome lost its dry-run mark: %s", raw)
+	}
+	raw, _ = json.Marshal(link.BuildResult{ID: "build-2"})
+	if string(raw) != `{"id":"build-2","repository":"","on":""}` {
+		t.Fatalf("an ordinary outcome carries a dry-run mark: %s", raw)
+	}
+}

cmd/mesh-controller/held_back.go

@@ -0,0 +1,241 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"sort"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// A push sends no build a policy or a plan holds back, except to the machine it names (novox/hq
+// issue 259, ADR 0221).
+//
+// A named push ends by sending every other machine whose declaration differs from what it was last
+// sent (ADR 0083), so that a grant the push's work minted reaches the provider in the same act. A
+// digest cannot say why a machine differs. A module whose upgrade policy records rather than rolls
+// out makes every machine running it differ from the merge on, and so did a module whose plan was
+// still waiting on its first machine (ADR 0218): `push <anchor>` sent all four machines the build
+// that was meant to be walked through the mesh one machine at a time, and a fault in it was met
+// everywhere at once.
+//
+// What tells the two apart is which build of each module the machine was last sent, kept with every
+// send. A machine any of whose modules would move to a build its policy or an open plan holds back
+// is not sent by a push that did not name it; the push says which, and why, and how to send it.
+
+// heldBack is why a push that did not name a machine must not send it, empty when it may.
+//
+// `modules` is what the machine would be sent now; `sent` and `known` what it was last sent, as
+// Inventory.SentBuilds answers. A module moves when the build it would carry is not the one the
+// machine was last sent — including a module the machine was never sent at all. A move is held when
+// the module's policy records rather than rolls out, or when an open plan has not yet sent this
+// machine (planStillToSend). A machine whose last send was not recorded is held whole: what it carried
+// is not known, so a held upgrade cannot be told from anything else.
+func heldBack(node string, modules []string, sent map[string]string, known bool,
+	current map[string]inventory.CurrentBuild, plans []inventory.Plan) []string {
+	if !known {
+		return []string{"which builds it was last sent is not known — it was last sent before the " +
+			"mesh kept them, or sent a declaration by hand"}
+	}
+	var why []string
+	for _, m := range modules {
+		now := current[m]
+		was, carried := sent[m]
+		if carried && was == now.Commit {
+			continue
+		}
+		move := fmt.Sprintf("%s would move %sto %s", m, fromBuild(was, carried), buildName(now.Commit))
+		if !now.RollOut {
+			why = append(why, move+", which its upgrade policy records rather than rolls out")
+			continue
+		}
+		if id := planStillToSend(plans, m, node); id != "" {
+			why = append(why, move+", which "+id+" has not sent it yet (one machine first)")
+		}
+	}
+	sort.Strings(why)
+	return why
+}
+
+// planStillToSend is the open plan that has a module's new build still to send this machine, or empty:
+// one holding the module that has neither finished sending it nor sent it here first, and has not
+// failed it (novox/hq ADR 0218). The same reading rolledOutByAPlan makes for the whole module, made
+// per machine.
+func planStillToSend(plans []inventory.Plan, module, node string) string {
+	for _, p := range plans {
+		s, holds := p.Modules[module]
+		if !p.Open() || !holds {
+			continue
+		}
+		if s == nil {
+			return p.ID
+		}
+		if s.SentAt != nil || s.State == "failed" {
+			continue
+		}
+		first := false
+		for _, n := range s.First {
+			if n == node {
+				first = true
+			}
+		}
+		if !first {
+			return p.ID
+		}
+	}
+	return ""
+}
+
+func fromBuild(was string, carried bool) string {
+	if !carried {
+		return "(never sent it) "
+	}
+	return "from " + buildName(was) + " "
+}
+
+func buildName(commit string) string {
+	if commit == "" {
+		return "a build with no source"
+	}
+	return shortCommit(commit)
+}
+
+// heldMachines reads, for each machine named, why a push that did not name it must not send it
+// (heldBack), and answers only the machines held. A machine whose set cannot be worked out is left
+// to the send, which says why.
+func heldMachines(ctx context.Context, open *stores, names []string) (map[string][]string, error) {
+	out := map[string][]string{}
+	if len(names) == 0 {
+		return out, nil
+	}
+	inv := open.inventory
+	current, err := inv.CurrentBuilds(ctx)
+	if err != nil {
+		return nil, err
+	}
+	plans, err := inv.OpenPlans(ctx)
+	if err != nil {
+		return nil, err
+	}
+	for _, node := range names {
+		plan, _, err := planFor(ctx, open, node)
+		if err != nil {
+			continue
+		}
+		modules := make([]string, 0, len(plan.Modules))
+		for _, m := range plan.Modules {
+			modules = append(modules, m.Module)
+		}
+		sent, known, err := inv.SentBuilds(ctx, node)
+		if err != nil {
+			return nil, err
+		}
+		if why := heldBack(node, modules, sent, known, current, plans); len(why) > 0 {
+			out[node] = why
+		}
+	}
+	return out, nil
+}
+
+// sayHeld is what a push says about a machine it left behind on purpose: that it is behind, why it
+// was not sent, that whatever else it is owed waits with it, and the command that sends it.
+func sayHeld(w io.Writer, node string, why []string) {
+	fmt.Fprintf(w, "\n%s is behind and was not sent: %s. A push sends no build a policy or a plan "+
+		"holds back to a machine it did not name (novox/hq ADR 0221), so anything else it is owed — a "+
+		"grant from this push among it — waits with it. `push %s` sends it\n",
+		node, strings.Join(why, "; "), node)
+}
+
+// flushBehind is the end of a named push: every other machine now behind is sent too, by name, over
+// as many rounds as the sends take to settle (novox/hq issue 057, ADR 0083) — except a machine whose
+// modules would move to a build a policy or a plan holds back, which is named and left (ADR 0221).
+//
+// `handled` is every machine already sent or already said; it is not considered again. Answers the
+// machines that could not be composed, as refusals.
+func flushBehind(ctx context.Context, open *stores, nodes []inventory.Node, handled map[string]bool,
+	compose func(held context.Context, node string) (sendable, error), d delivery, holder string,
+	w io.Writer) ([]string, error) {
+	inv := open.inventory
+	var refusals []string
+	// Bounded by the node count: a node is marked handled the round it is considered and is never
+	// considered twice, so the loop cannot run more than len(nodes) rounds. The bound is a guard
+	// against a logic error, not a real limit — if it were ever hit, that is a bug rather than a
+	// cascade legitimately still converging, so it is said rather than passed over in silence.
+	rounds := 0
+	for {
+		would, err := wouldSend(ctx, open, nodes)
+		if err != nil {
+			return refusals, err
+		}
+		behind, err := inv.Waiting(ctx, would)
+		if err != nil {
+			return refusals, err
+		}
+		var also []string
+		for _, m := range behind {
+			if !handled[m.Node] {
+				also = append(also, m.Node)
+			}
+		}
+		if len(also) == 0 {
+			return refusals, nil
+		}
+		if rounds++; rounds > len(nodes) {
+			fmt.Fprintf(w, "\nstopped cascading after %d rounds with %s still behind — this "+
+				"should not happen; run `push --behind` to finish\n",
+				rounds-1, strings.Join(also, ", "))
+			return refusals, nil
+		}
+		sort.Strings(also)
+		held, err := heldMachines(ctx, open, also)
+		if err != nil {
+			return refusals, err
+		}
+		var sending []string
+		for _, name := range also {
+			// Every candidate this round is marked handled — the sent ones so they are not
+			// re-listed, the held ones because they stay held, and the refused ones so a machine
+			// that cannot be composed does not make the loop spin on it for ever.
+			handled[name] = true
+			if why, isHeld := held[name]; isHeld {
+				sayHeld(w, name, why)
+				continue
+			}
+			sending = append(sending, name)
+		}
+		if len(sending) == 0 {
+			continue
+		}
+		fmt.Fprintf(w, "\nthis push left %s behind — a provision granted from there, or a "+
+			"declaration since changed; sending it too\n", strings.Join(sending, ", "))
+		// Tolerantly, exactly as the named send: a machine that cannot be composed is collected as
+		// a refusal and reported at the end, and the others are still sent (novox/hq ADR 0066).
+		// Held for this round only, and after the last round's were given back, so two pushes
+		// cascading into each other's machines never each wait on the other.
+		refused, err := sendRound(ctx, open, sending, compose, d, holder)
+		refusals = append(refusals, refused...)
+		if err != nil {
+			return refusals, err
+		}
+	}
+}
+
+// composeForPush is how a push composes one machine: its set resolved, what it cannot host and what
+// is left out of it said, and its declaration allocated.
+func composeForPush(open *stores, gens map[string]catalogue.Generator) func(held context.Context, node string) (sendable, error) {
+	return func(held context.Context, node string) (sendable, error) {
+		plan, settings, err := planFor(held, open, node)
+		if err != nil {
+			return sendable{}, err
+		}
+		reportUnhostable(node, plan)
+		declared, err := declarationWith(held, open, node, plan, settings, gens, Allocating)
+		if err == nil {
+			reportLeftOut(node, declared)
+		}
+		return declared, err
+	}
+}

cmd/mesh-controller/held_back_test.go

@@ -0,0 +1,335 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"reflect"
+	"slices"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/overlay"
+)
+
+// novox/hq issue 259, ADR 0221: a push that did not name a machine does not send it a build its
+// upgrade policy records rather than rolls out, nor one an open plan has not sent it yet. Anything
+// else that moved is still a consequence the push sends (ADR 0083).
+func TestAHeldBuildHoldsAMachineANamedPushDidNotName(t *testing.T) {
+	current := map[string]inventory.CurrentBuild{
+		"resolver": {Commit: "c2c2c2c2c2"},
+		"agent":    {Commit: "a2", RollOut: true},
+		"network":  {},
+	}
+	modules := []string{"network", "resolver", "agent"}
+	sent := map[string]string{"network": "", "resolver": "c1c1c1c1c1", "agent": "a2"}
+
+	// A module whose policy records moved: held, naming it, both builds and why.
+	why := heldBack("laptop", modules, sent, true, current, nil)
+	if len(why) != 1 || !strings.Contains(why[0], "resolver would move from c1c1c1c1 to c2c2c2c2") ||
+		!strings.Contains(why[0], "upgrade policy records") {
+		t.Fatalf("a recorded upgrade did not hold the machine: %v", why)
+	}
+
+	// Nothing moved — what differs is a grant, a peer, a setting: not held (issue 057).
+	sent["resolver"] = "c2c2c2c2c2"
+	if why := heldBack("laptop", modules, sent, true, current, nil); len(why) != 0 {
+		t.Fatalf("a machine whose builds are all current was held: %v", why)
+	}
+
+	// A module whose policy rolls out moved, and no plan holds it: sent, as before.
+	sent["agent"] = "a1"
+	if why := heldBack("laptop", modules, sent, true, current, nil); len(why) != 0 {
+		t.Fatalf("a rolled-out upgrade no plan holds was held: %v", why)
+	}
+
+	// The last send's builds are not known: held whole.
+	if why := heldBack("laptop", modules, nil, false, current, nil); len(why) != 1 ||
+		!strings.Contains(why[0], "not known") {
+		t.Fatalf("a machine whose last send was not recorded was not held: %v", why)
+	}
+
+	// A module the machine was never sent, under a recording policy: held, and said so.
+	delete(sent, "resolver")
+	sent["agent"] = "a2"
+	if why := heldBack("laptop", modules, sent, true, current, nil); len(why) != 1 ||
+		!strings.Contains(why[0], "resolver would move (never sent it) to c2c2c2c2") {
+		t.Fatalf("a module never sent under a recording policy: %v", why)
+	}
+}
+
+// ADR 0218 meets ADR 0083: a plan waiting on its first machine has not sent the rest, and a push
+// naming some other machine must not send them for it.
+func TestAPlanWaitingOnItsFirstMachineHoldsTheRest(t *testing.T) {
+	at := time.Now()
+	current := map[string]inventory.CurrentBuild{"agent": {Commit: "a2", RollOut: true}}
+	sent := map[string]string{"agent": "a1"}
+	waiting := []inventory.Plan{{ID: "plan-7", State: inventory.PlanRolling, Modules: map[string]*inventory.PlanModule{
+		"agent": {State: "built", First: []string{"ace"}, FirstAt: &at}}}}
+
+	why := heldBack("g14", []string{"agent"}, sent, true, current, waiting)
+	if len(why) != 1 || !strings.Contains(why[0], "plan-7 has not sent it yet") {
+		t.Fatalf("a machine the plan has not reached was not held: %v", why)
+	}
+	// The first machine itself was sent by the plan: not held by it.
+	if why := heldBack("ace", []string{"agent"}, sent, true, current, waiting); len(why) != 0 {
+		t.Fatalf("the plan's first machine was held: %v", why)
+	}
+	// Built but not yet sent anywhere, or not yet built: the plan has it still to send.
+	for what, s := range map[string]*inventory.PlanModule{"built, unsent": {State: "built"}, "unasked": nil} {
+		plans := []inventory.Plan{{ID: "plan-8", State: inventory.PlanBuilding,
+			Modules: map[string]*inventory.PlanModule{"agent": s}}}
+		if why := heldBack("ace", []string{"agent"}, sent, true, current, plans); len(why) != 1 {
+			t.Errorf("%s: not held: %v", what, why)
+		}
+	}
+	// Sent everywhere, failed, or a plan no longer open: the plan holds nothing back.
+	for what, plans := range map[string][]inventory.Plan{
+		"sent everywhere": {{ID: "p", State: inventory.PlanRolling, Modules: map[string]*inventory.PlanModule{
+			"agent": {State: "built", First: []string{"ace"}, FirstAt: &at, SentAt: &at}}}},
+		"failed": {{ID: "p", State: inventory.PlanRolling, Modules: map[string]*inventory.PlanModule{
+			"agent": {State: "failed"}}}},
+		"closed": {{ID: "p", State: inventory.PlanDone, Modules: map[string]*inventory.PlanModule{
+			"agent": {State: "built"}}}},
+	} {
+		if why := heldBack("g14", []string{"agent"}, sent, true, current, plans); len(why) != 0 {
+			t.Errorf("%s: held: %v", what, why)
+		}
+	}
+}
+
+// A send records the build of each module it carried; a module left out of it keeps the build it
+// was last sent, since the machine keeps that one.
+func TestASendCarriesTheCurrentBuildsAndALeftOutModuleKeepsItsOwn(t *testing.T) {
+	current := map[string]inventory.CurrentBuild{"a": {Commit: "a2"}, "b": {Commit: "b2"}, "c": {}}
+	got := carriedBuilds([]string{"a", "b", "c"}, map[string]string{"b": "a setting does not compose"},
+		current, map[string]string{"a": "a1", "b": "b1"})
+	if want := map[string]string{"a": "a2", "b": "b1", "c": ""}; !reflect.DeepEqual(got, want) {
+		t.Fatalf("carried %v, wanted %v", got, want)
+	}
+	// Not known before: the left-out module is not recorded at all, so it reads as never sent.
+	got = carriedBuilds([]string{"a", "b"}, map[string]string{"b": "x"}, current, nil)
+	if want := map[string]string{"a": "a2"}; !reflect.DeepEqual(got, want) {
+		t.Fatalf("carried %v, wanted %v", got, want)
+	}
+}
+
+// recordedDelivery sends nothing and records each send as the mesh does, so the next comparison
+// reads the machine as current — and writes down which machines it declared.
+type recordedDelivery struct {
+	inv      *inventory.Inventory
+	declared []string
+}
+
+func (r *recordedDelivery) grant(context.Context, []readyNode) error { return nil }
+
+func (r *recordedDelivery) declare(ctx context.Context, s readyNode, body []byte) (string, error) {
+	r.declared = append(r.declared, s.node)
+	return recordSent(ctx, r.inv, s.node, body, s.declared.Builds)
+}
+
+// aResolver is a module built from a repository, at a commit, with something on the machine that
+// says which build it is.
+func aResolver(t *testing.T, open *stores, commit string, asked time.Time) {
+	t.Helper()
+	m := catalogue.Manifest{Module: "resolver", Version: "1", Resources: []map[string]any{
+		{"id": "zones", "type": "file", "path": "/etc/resolver/zones", "content": "built from " + commit},
+	}}
+	if err := open.inventory.RegisterModule(t.Context(), m, inventory.Source{
+		Repository: "novox/mesh-catalog", Path: "modules/resolver", BuiltFrom: commit, Asked: asked}); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// A third machine on the private network: once it is sent, every other machine's peers change with
+// it, which is a consequence a push must still send — no build moved.
+func aThirdMachine(t *testing.T, open *stores) {
+	t.Helper()
+	ctx := t.Context()
+	record, err := open.inventory.AddNode(ctx, "spare")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.SetPlace(ctx, "spare", "spare.example:51820", "here", false, "10.77.0.3"); err != nil {
+		t.Fatal(err)
+	}
+	reported, err := json.Marshal(map[string]any{"capabilities": []map[string]any{
+		{"name": "container-runtime", "present": true}, {"name": "wireguard", "present": true},
+		{"name": "systemd", "present": true}}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var profile map[string]any
+	if err := json.Unmarshal(reported, &profile); err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.RecordProfile(ctx, record.ID, profile); err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.RecordSealingKey(ctx, record.ID, aPublicKey(t)); err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.RecordOverlayKey(ctx, record.ID, aPublicKey(t)); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := open.inventory.Assign(ctx, "spare", overlay.Name); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// The issue as it happened, against the real stores: a change merged with the policy `record`, a push
+// naming the anchor, and the laptop — running the same module — left with what it had, by name.
+func TestANamedPushLeavesAMachineAPolicyHoldsBack(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+	asked := time.Now().Add(-time.Hour)
+	aResolver(t, open, "c1c1c1c1c1", asked)
+	for _, node := range []string{"anchor", "laptop"} {
+		if _, err := inv.Assign(ctx, node, "resolver"); err != nil {
+			t.Fatal(err)
+		}
+	}
+	gens, err := generators(ctx, open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	compose := composeForPush(open, gens)
+	d := &recordedDelivery{inv: inv}
+	if _, err := sendRound(ctx, open, []string{"anchor", "laptop"}, compose, d, ""); err != nil {
+		t.Fatal(err)
+	}
+	if builds, known, err := inv.SentBuilds(ctx, "laptop"); err != nil || !known || builds["resolver"] != "c1c1c1c1c1" {
+		t.Fatalf("the send did not record the build it carried: %v %v %v", builds, known, err)
+	}
+	digestOfLaptop := func() string {
+		sent, err := inv.Outstanding(ctx, "laptop")
+		if err != nil {
+			t.Fatal(err)
+		}
+		return sent
+	}
+	before := digestOfLaptop()
+
+	// The change merges; the policy is the default, record. `push anchor` sends the anchor...
+	aResolver(t, open, "c2c2c2c2c2", asked.Add(time.Minute))
+	d.declared = nil
+	if _, err := sendRound(ctx, open, []string{"anchor"}, compose, d, ""); err != nil {
+		t.Fatal(err)
+	}
+	// ...and its cascade leaves the laptop, saying so.
+	var said bytes.Buffer
+	d.declared = nil
+	refused, err := flushBehind(ctx, open, mustNodes(t, open), map[string]bool{"anchor": true}, compose, d, "", &said)
+	if err != nil || len(refused) != 0 {
+		t.Fatalf("the cascade failed: %v %v", refused, err)
+	}
+	if len(d.declared) != 0 {
+		t.Fatalf("the cascade sent %v a build its policy records", d.declared)
+	}
+	if digestOfLaptop() != before {
+		t.Fatal("the laptop's last send moved: it was sent the held build")
+	}
+	for _, want := range []string{"laptop is behind and was not sent", "resolver would move from c1c1c1c1 to c2c2c2c2",
+		"upgrade policy records", "`push laptop` sends it"} {
+		if !strings.Contains(said.String(), want) {
+			t.Errorf("the push did not say %q:\n%s", want, said.String())
+		}
+	}
+
+	// Held and owed something else at once — a peer joined: still not sent, and both said: why it
+	// is held, and that what else it is owed waits with it.
+	aThirdMachine(t, open)
+	d.declared = nil
+	if _, err := sendRound(ctx, open, []string{"spare"}, compose, d, ""); err != nil {
+		t.Fatal(err)
+	}
+	d.declared = nil
+	said.Reset()
+	if _, err := flushBehind(ctx, open, mustNodes(t, open), map[string]bool{"anchor": true, "spare": true},
+		compose, d, "", &said); err != nil {
+		t.Fatal(err)
+	}
+	if len(d.declared) != 0 || digestOfLaptop() != before {
+		t.Fatalf("a held machine owed a consequence was sent: %v", d.declared)
+	}
+	if !strings.Contains(said.String(), "resolver would move") || !strings.Contains(said.String(), "anything else it is owed") {
+		t.Fatalf("the push did not say both:\n%s", said.String())
+	}
+
+	// A policy that rolls out: the laptop is a consequence like any other, and sent.
+	if err := inv.SetUpgradeOf(ctx, "resolver", inventory.Upgrade{RollOut: true}); err != nil {
+		t.Fatal(err)
+	}
+	said.Reset()
+	if _, err := flushBehind(ctx, open, mustNodes(t, open), map[string]bool{"anchor": true, "spare": true},
+		compose, d, "", &said); err != nil {
+		t.Fatal(err)
+	}
+	if !reflect.DeepEqual(d.declared, []string{"laptop"}) || digestOfLaptop() == before {
+		t.Fatalf("a rolled-out upgrade's machine was not sent: %v\n%s", d.declared, said.String())
+	}
+	if builds, _, _ := inv.SentBuilds(ctx, "laptop"); builds["resolver"] != "c2c2c2c2c2" {
+		t.Fatalf("the new send did not record the new build: %v", builds)
+	}
+}
+
+// Issue 057's case is unchanged: a machine whose builds are all current and whose declaration moved
+// for another reason is sent by a push that names someone else.
+func TestANamedPushStillSendsAConsequenceNothingHolds(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+	aResolver(t, open, "c1c1c1c1c1", time.Now().Add(-time.Hour))
+	if _, err := inv.Assign(ctx, "laptop", "resolver"); err != nil {
+		t.Fatal(err)
+	}
+	gens, err := generators(ctx, open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	compose := composeForPush(open, gens)
+	d := &recordedDelivery{inv: inv}
+	if _, err := sendRound(ctx, open, []string{"anchor", "laptop"}, compose, d, ""); err != nil {
+		t.Fatal(err)
+	}
+
+	// `push spare`, the machine just placed: the others' peers change with it.
+	aThirdMachine(t, open)
+	if _, err := sendRound(ctx, open, []string{"spare"}, compose, d, ""); err != nil {
+		t.Fatal(err)
+	}
+	d.declared = nil
+	var said bytes.Buffer
+	if _, err := flushBehind(ctx, open, mustNodes(t, open), map[string]bool{"spare": true}, compose, d, "", &said); err != nil {
+		t.Fatal(err)
+	}
+	if !reflect.DeepEqual(d.declared, []string{"anchor", "laptop"}) {
+		t.Fatalf("a consequence nothing holds was not sent: %v\n%s", d.declared, said.String())
+	}
+	if strings.Contains(said.String(), "was not sent") {
+		t.Fatalf("a machine nothing holds was said to be held:\n%s", said.String())
+	}
+
+	// A machine whose last send was not recorded — a declaration sent by hand — is held until named.
+	record, err := inv.NodeByName(ctx, "laptop")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := inv.RecordSent(ctx, record.ID, "sent-by-hand", nil); err != nil {
+		t.Fatal(err)
+	}
+	d.declared = nil
+	said.Reset()
+	if _, err := flushBehind(ctx, open, mustNodes(t, open), map[string]bool{"spare": true}, compose, d, "", &said); err != nil {
+		t.Fatal(err)
+	}
+	// The anchor, the hub, may still be settling from the machine placed above; the laptop is the
+	// question.
+	if slices.Contains(d.declared, "laptop") || !strings.Contains(said.String(), "laptop is behind and was not sent") {
+		t.Fatalf("a machine whose last send is not known was sent: %v\n%s", d.declared, said.String())
+	}
+}

cmd/mesh-controller/hold.go

@@ -0,0 +1,37 @@
+package main
+
+import (
+	"context"
+)
+
+// heldKey carries the nodes this call already holds, so an act that holds a node and then sends
+// through sendTo does not wait on itself.
+type heldKey struct{}
+
+// holdNodes holds the named nodes for composing and sending their declarations (novox/hq ADR
+// 0100), skipping any the context already holds, and returns a context that says it holds them.
+// Release gives back only what this call took.
+func holdNodes(ctx context.Context, open *stores, names []string) (context.Context, func(), error) {
+	already, _ := ctx.Value(heldKey{}).(map[string]bool)
+	var take []string
+	for _, n := range names {
+		if !already[n] {
+			take = append(take, n)
+		}
+	}
+	if len(take) == 0 {
+		return ctx, func() {}, nil
+	}
+	release, err := open.inventory.HoldNodes(ctx, take)
+	if err != nil {
+		return ctx, nil, err
+	}
+	held := map[string]bool{}
+	for n := range already {
+		held[n] = true
+	}
+	for _, n := range take {
+		held[n] = true
+	}
+	return context.WithValue(ctx, heldKey{}, held), release, nil
+}

cmd/mesh-controller/hold_test.go

@@ -0,0 +1,45 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+)
+
+// A cascade round gives its hold back on every way out: a declaration that cannot be marshalled
+// and a send that fails leave nobody waiting for the node.
+func TestASendRoundGivesItsHoldBackOnEveryWayOut(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	unmarshallable := func(context.Context, string) (sendable, error) {
+		return sendable{Resources: []map[string]any{{"id": "x", "bad": make(chan int)}}}, nil
+	}
+	plain := func(context.Context, string) (sendable, error) {
+		return sendable{Resources: []map[string]any{{"id": "x"}}}, nil
+	}
+	failing := &recordingDelivery{declareErr: errors.New("the broker went away")}
+	fine := &recordingDelivery{}
+
+	for name, round := range map[string]func() error{
+		"a body that cannot be marshalled": func() error {
+			_, err := sendRound(ctx, open, []string{"anchor"}, unmarshallable, fine, "")
+			return err
+		},
+		"a send that fails": func() error {
+			_, err := sendRound(ctx, open, []string{"anchor"}, plain, failing, "")
+			return err
+		},
+	} {
+		if err := round(); err == nil {
+			t.Fatalf("%s was not an error", name)
+		}
+		waiting, cancel := context.WithTimeout(ctx, 2*time.Second)
+		release, err := open.inventory.HoldNodes(waiting, []string{"anchor"})
+		cancel()
+		if err != nil {
+			t.Fatalf("after %s the node is still held: %v", name, err)
+		}
+		release()
+	}
+}

cmd/mesh-controller/holdings.go

@@ -0,0 +1,163 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"sort"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/overlay"
+)
+
+// recordDerivedHolders writes down who holds each mesh-scoped seat that nobody was ever recorded
+// as holding.
+//
+// **A seat held by derivation is a seat held by accident of being alone** (novox/hq
+// 04-ISSUES/170). ADR 0131 lets a holder on record settle a seat, and lets any other assignment
+// whose module could hold it stand beside the holder, eligible and silent. But a seat nobody
+// ever handed over has no record, so its holder is whichever assignment happened to be the sole
+// claimant — and the day a second one is assigned, both claim, both are refused, and the first
+// one's whole machine stops resolving. That is what assigning a second postgres did to the
+// control plane's own store.
+//
+// So the mesh writes the derived answer down before it acts on an assignment: for every
+// mesh-scoped seat with exactly one resolved holder and nothing on record, that holder is
+// recorded as the standing one — the same record `seat <name> --to <node>/<module>` makes by
+// hand, made from what the mesh already resolved. A seat with two derived claimants is left
+// alone: that is the ambiguity a person settles, and recording either would be guessing.
+//
+// Node-scoped seats are untouched: a record is one holder per seat, and a node-scoped seat has
+// one holder per machine (ADR 0121), so there is nothing for a record to settle there.
+func recordDerivedHolders(ctx context.Context, open *stores) ([]string, error) {
+	inv := open.inventory
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return nil, err
+	}
+	// exclude nobody: every node's claims, resolved with the holdings on record.
+	world, err := theRestOfTheMesh(ctx, inv, shelf, "")
+	if err != nil {
+		return nil, err
+	}
+	recorded, err := inv.Holdings(ctx)
+	if err != nil {
+		return nil, err
+	}
+	// A record is a row against a seat the store knows. A seat it does not — a mesh whose seats
+	// were never seeded, a seat a module declares for itself — stays held by derivation, as it
+	// always was; a missing row is not a reason an assignment fails.
+	known, err := inv.Seats(ctx)
+	if err != nil {
+		return nil, err
+	}
+	recordable := map[string]bool{}
+	for _, s := range known {
+		recordable[s.Name] = true
+	}
+	onRecord := map[string]bool{}
+	for _, h := range recorded {
+		if s, ok := catalogue.SeatNamed(h.Claim); ok {
+			onRecord[s.Name] = true
+		}
+	}
+	holders := map[string][]catalogue.Held{}
+	for _, h := range world.Held {
+		if h.Scope != catalogue.ScopeMesh {
+			continue
+		}
+		s, ok := catalogue.SeatNamed(h.Claim)
+		if !ok || onRecord[s.Name] || !recordable[s.Name] {
+			continue
+		}
+		holders[s.Name] = append(holders[s.Name], h)
+	}
+	names := make([]string, 0, len(holders))
+	for name := range holders {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	var said []string
+	for _, name := range names {
+		if len(holders[name]) != 1 {
+			continue
+		}
+		h := holders[name][0]
+		if err := inv.HoldSeat(ctx, name, catalogue.ScopeMesh, h.Node, h.Module); err != nil {
+			return said, err
+		}
+		said = append(said, fmt.Sprintf(
+			"recorded %s on %s as the standing holder of %s, which it held only by being alone",
+			h.Module, h.Node, name))
+	}
+	return said, nil
+}
+
+// replicatedHolders is, for each replicated mesh seat, every machine on the private network holding
+// it — by internal name, at its private address (novox/hq ADR 0223). What a machine's resolver file
+// lists for `mesh-dns-resolver`; the rendering puts the machine itself first when it is one.
+//
+// **The holders on record, and only the sole claimant when there are none** — the same answer the
+// resolver gives about who holds (ADR 0131, issue 170). An assignment standing beside the holders,
+// eligible and silent, is not listed: it becomes a holder by `seat <name> --add`, an act, never by
+// being assigned. Two claimants with nothing on record are refused at resolution, so neither is
+// listed here. A holder off the private network is left out: a resolver named at an address nothing
+// answers is a lookup that waits out its timeout on every name.
+func replicatedHolders(ctx context.Context, inv *inventory.Inventory,
+	shelf map[string]catalogue.Manifest) (map[string]map[string]string, error) {
+	var replicated []catalogue.Seat
+	for _, s := range catalogue.Seats() {
+		if s.Replicated && s.Scope == catalogue.ScopeMesh {
+			replicated = append(replicated, s)
+		}
+	}
+	if len(replicated) == 0 {
+		return nil, nil
+	}
+	recorded, err := inv.Holdings(ctx)
+	if err != nil {
+		return nil, err
+	}
+	places, err := onTheNetwork(ctx, inv, shelf)
+	if err != nil {
+		return nil, err
+	}
+	address := map[string]string{}
+	for _, p := range places {
+		address[p.Name] = p.Address
+	}
+	entries, err := inv.Catalogued(ctx)
+	if err != nil {
+		return nil, err
+	}
+	out := map[string]map[string]string{}
+	for _, seat := range replicated {
+		var nodes []string
+		for _, h := range recorded {
+			if hs, ok := catalogue.SeatNamed(h.Claim); ok && hs.Name == seat.Name && h.Scope == seat.Scope {
+				nodes = append(nodes, h.Node)
+			}
+		}
+		if len(nodes) == 0 {
+			var derived []string
+			for _, e := range entries {
+				for _, c := range e.Manifest.Claims {
+					if cs, ok := catalogue.SeatNamed(c.Name); ok && cs.Name == seat.Name && c.At() == seat.Scope {
+						derived = append(derived, e.On...)
+					}
+				}
+			}
+			if len(derived) == 1 {
+				nodes = derived
+			}
+		}
+		at := map[string]string{}
+		for _, n := range nodes {
+			if address[n] != "" {
+				at[overlay.InternalName(n)] = address[n]
+			}
+		}
+		out[seat.Name] = at
+	}
+	return out, nil
+}

cmd/mesh-controller/holdings_test.go

@@ -0,0 +1,110 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// A seat nobody ever handed over is held by whichever assignment happened to be alone — and the
+// day a second module able to hold it is assigned, both claimed, both were refused, and the first
+// one's machine stopped resolving (novox/hq 04-ISSUES/170). The mesh now writes the derived holder
+// down before it acts, so the second assignment stands beside the holder on record.
+
+func aSeatedStore() catalogue.Manifest {
+	return catalogue.Manifest{Module: "store", Version: "1",
+		Provides: []catalogue.Offer{{Name: "postgres-database", Scope: catalogue.ScopeMesh}},
+		Serves:   map[string]map[string]any{"postgres-database": {"port": 5432}},
+		Claims:   []catalogue.Claim{{Name: "mesh-store", Scope: catalogue.ScopeMesh}}}
+}
+
+func TestASecondEligibleHolderStandsBesideTheOneHeldByBeingAlone(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	// Every deploy seeds the mesh's own seats; a record is a row against one of them.
+	if _, err := open.inventory.SeedSeats(ctx, catalogue.DefaultSeats()); err != nil {
+		t.Fatal(err)
+	}
+	register(t, open, aSeatedStore())
+
+	if _, err := assign(ctx, open, "anchor", "store"); err != nil {
+		t.Fatal(err)
+	}
+	said, err := assign(ctx, open, "laptop", "store")
+	if err != nil {
+		t.Fatalf("a second store, eligible for the seat, was refused:\n%s\n%v", said, err)
+	}
+	if strings.Contains(said, "cannot be worked out") {
+		t.Fatalf("assigning a second store unsettled the first one's machine:\n%s", said)
+	}
+	if !strings.Contains(said, "recorded store on anchor as the standing holder of mesh-store") {
+		t.Fatalf("the holder by derivation was not written down:\n%s", said)
+	}
+
+	holdings, err := open.inventory.Holdings(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var found bool
+	for _, h := range holdings {
+		if h.Claim == "mesh-store" {
+			found = true
+			if h.Node != "anchor" || h.Module != "store" {
+				t.Fatalf("mesh-store is recorded on %s/%s, not on the one that held it", h.Node, h.Module)
+			}
+		}
+	}
+	if !found {
+		t.Fatalf("mesh-store has no holder on record after assigning: %v", holdings)
+	}
+
+	// And the record decides from here: the anchor's plan holds the seat, the laptop's does not.
+	for node, holds := range map[string]bool{"anchor": true, "laptop": false} {
+		plan, _, err := planFor(ctx, open, node)
+		if err != nil {
+			t.Fatalf("%s no longer resolves: %v", node, err)
+		}
+		var claimed bool
+		for _, c := range plan.Claims {
+			if c.Claim == "mesh-store" {
+				claimed = true
+			}
+		}
+		if claimed != holds {
+			t.Fatalf("%s holds mesh-store: %v, want %v", node, claimed, holds)
+		}
+	}
+}
+
+func TestAHolderOnRecordIsNotRewrittenByDerivation(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	if _, err := open.inventory.SeedSeats(ctx, catalogue.DefaultSeats()); err != nil {
+		t.Fatal(err)
+	}
+	register(t, open, aSeatedStore())
+	for _, node := range []string{"anchor", "laptop"} {
+		if _, err := assign(ctx, open, node, "store"); err != nil {
+			t.Fatal(err)
+		}
+	}
+	// A person hands the seat to the laptop. From here the record decides, and what the mesh
+	// derives must never write over it.
+	if err := open.inventory.HoldSeat(ctx, "mesh-store", catalogue.ScopeMesh, "laptop", "store"); err != nil {
+		t.Fatal(err)
+	}
+	said, err := recordDerivedHolders(ctx, open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(said) != 0 {
+		t.Fatalf("a seat on record was written again from derivation: %v", said)
+	}
+	holdings, _ := open.inventory.Holdings(ctx)
+	for _, h := range holdings {
+		if h.Claim == "mesh-store" && h.Node != "laptop" {
+			t.Fatalf("the record moved to %s", h.Node)
+		}
+	}
+}

cmd/mesh-controller/hosts_behind_test.go

@@ -0,0 +1,111 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// A host refuses a declaration carrying a field it does not know, and refuses it whole — so every new
+// field is a flag day, and the mesh had no record of which host any machine ran (novox/hq
+// 04-ISSUES/087). The order was kept by somebody remembering it.
+
+func TestTheMeshNamesWhichMachinesRunWhichHost(t *testing.T) {
+	split := hostSplit([]inventory.Node{
+		{Name: "anchor", HostVersion: "04a27ca"},
+		{Name: "laptop", HostVersion: "ced54d4"},
+		{Name: "spare", HostVersion: "04a27ca"},
+	})
+	if len(split) != 2 {
+		t.Fatalf("two versions were reported and the split has %d: %v", len(split), split)
+	}
+	if got := strings.Join(split["04a27ca"], ","); got != "anchor,spare" && got != "spare,anchor" {
+		t.Fatalf("04a27ca is held by %q", got)
+	}
+	if got := strings.Join(split["ced54d4"], ","); got != "laptop" {
+		t.Fatalf("ced54d4 is held by %q", got)
+	}
+}
+
+func TestTheMeshDoesNotClaimWhichHostIsNewer(t *testing.T) {
+	// **The fault this replaced.** A host reports its version as a commit, and commits have no order.
+	// The first version compared them as strings and, on the live mesh, named the three machines
+	// running the NEWER host as the ones behind: `ced54d4` sorts above `04a27ca` and means nothing.
+	//
+	// There is no assertion to make about which is newer, and that is the point — the type says so.
+	// hostSplit returns who runs what, and nothing that could be read as an ordering.
+	split := hostSplit([]inventory.Node{
+		{Name: "old-but-sorts-high", HostVersion: "ced54d4"},
+		{Name: "new-but-sorts-low", HostVersion: "04a27ca"},
+	})
+	for version, machines := range split {
+		if len(machines) != 1 {
+			t.Fatalf("%s is held by %v", version, machines)
+		}
+	}
+}
+
+func TestAMachineThatHasNotSaidIsNotAVersion(t *testing.T) {
+	// It may be running anything. Counting it as a version would invent a disagreement; `node show`
+	// says per machine that it has not said.
+	split := hostSplit([]inventory.Node{
+		{Name: "anchor", HostVersion: "04a27ca"},
+		{Name: "quiet"},
+	})
+	if split != nil {
+		t.Fatalf("one reported version and one silence read as a disagreement: %v", split)
+	}
+}
+
+func TestMachinesAgreeingOnTheirHostAreNotADisagreement(t *testing.T) {
+	if split := hostSplit([]inventory.Node{
+		{Name: "anchor", HostVersion: "v2"},
+		{Name: "laptop", HostVersion: "v2"},
+	}); split != nil {
+		t.Fatalf("machines agreeing reported a split: %v", split)
+	}
+}
+
+func TestAMeshWhereNothingReportedAHostStatesNoDisagreement(t *testing.T) {
+	if split := hostSplit([]inventory.Node{{Name: "anchor"}, {Name: "laptop"}}); split != nil {
+		t.Fatalf("a mesh told no host version reported a split: %v", split)
+	}
+}
+
+func TestAReportedHostVersionIsKeptAndReadBack(t *testing.T) {
+	// The machine has sent this since ADR 0141 and the controller's own copy of the report did not
+	// have the field, so it was unmarshalled into nothing. End to end through the store, because the
+	// fault was a field that existed on one side of the wire only.
+	open := aMesh(t)
+	record, err := open.inventory.NodeByName(t.Context(), "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if record.HostVersion != "" {
+		t.Fatalf("a machine that never reported one has host version %q", record.HostVersion)
+	}
+	if err := open.inventory.RecordHostVersion(t.Context(), record.ID, "ced54d4"); err != nil {
+		t.Fatal(err)
+	}
+	again, err := open.inventory.NodeByName(t.Context(), "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if again.HostVersion != "ced54d4" {
+		t.Fatalf("the reported host version read back as %q", again.HostVersion)
+	}
+	// An empty report never clears what a machine last said: a bare word that the node is there says
+	// nothing about its host.
+	if err := open.inventory.RecordHostVersion(t.Context(), record.ID, "  "); err != nil {
+		t.Fatal(err)
+	}
+	kept, err := open.inventory.NodeByName(t.Context(), "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if kept.HostVersion != "ced54d4" {
+		t.Fatalf("a report carrying no host version cleared what the machine had said: %q",
+			kept.HostVersion)
+	}
+}

cmd/mesh-controller/issue_204_test.go

@@ -0,0 +1,93 @@
+package main
+
+import (
+	"context"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// A declaration composed earlier is numbered lower than one composed later, whatever order the two
+// are sent in (novox/hq issue 204). The number used to be taken at send time, after composing, so a
+// declaration composed before an assignment changed and sent after a newer one carried the higher
+// number — and the machine, which refuses a lower number, took the older content as the mesh's
+// newest word. Taken before the composition reads anything, the order of numbers is the order of
+// compositions, and the host's refusal does what it is for.
+func TestADeclarationComposedEarlierIsNumberedLowerWhateverOrderItIsSent(t *testing.T) {
+	allot := numbered()
+	var composed []string
+	compose := func(stamp string) func(string) (sendable, error) {
+		return func(node string) (sendable, error) {
+			composed = append(composed, stamp)
+			return sendable{Resources: []map[string]any{{"id": node + "." + stamp}}}, nil
+		}
+	}
+	// Composed first — before an assignment changed — and sent last.
+	stale, _ := composeEach([]string{"anchor"}, allot, compose("before"))
+	// Composed after the change, sent first.
+	fresh, _ := composeEach([]string{"anchor"}, allot, compose("after"))
+
+	if stale[0].declared.Sequence != 1 || fresh[0].declared.Sequence != 2 {
+		t.Fatalf("the numbers do not follow the compositions: before=%d after=%d",
+			stale[0].declared.Sequence, fresh[0].declared.Sequence)
+	}
+	// Sent in the other order, the numbers do not change — so the machine that has applied the
+	// fresh one (2) refuses the stale one (1) when it arrives late.
+	if !(stale[0].declared.Sequence < fresh[0].declared.Sequence) {
+		t.Fatal("a declaration composed earlier must carry the lower number, however late it is sent")
+	}
+	if len(composed) != 2 || composed[0] != "before" {
+		t.Fatalf("compositions happened in an unexpected order: %v", composed)
+	}
+}
+
+// The number is taken before the first read of the composition, not after it: an allotter that
+// fails leaves nothing composed for that machine, and the others are still composed.
+func TestTheNumberIsTakenBeforeComposingAndItsFailureIsARefusal(t *testing.T) {
+	calls := 0
+	allot := func(node string) (int64, error) {
+		if node == "anchor" {
+			return 0, context.DeadlineExceeded
+		}
+		return 7, nil
+	}
+	sending, refusals := composeEach([]string{"anchor", "laptop"}, allot, func(node string) (sendable, error) {
+		calls++
+		if node == "anchor" {
+			t.Fatal("anchor was composed although its number could not be taken")
+		}
+		return sendable{}, nil
+	})
+	if calls != 1 || len(sending) != 1 || sending[0].node != "laptop" || sending[0].declared.Sequence != 7 {
+		t.Fatalf("laptop should be composed with its number and anchor refused: %v / %v", sending, refusals)
+	}
+	if len(refusals) != 1 {
+		t.Fatalf("anchor's failed number should be a refusal naming it: %v", refusals)
+	}
+}
+
+// What was sent is written down even when the sender's context is already cancelled (issue 204): a
+// controller replaced mid-send had told the machine and never recorded it, so status read "applied,
+// current" over a machine that had just been sent something else.
+func TestASendIsRecordedEvenWhenTheSenderIsBeingCancelled(t *testing.T) {
+	inv := inventory.ForTest(t)
+	ctx, cancel := context.WithCancel(t.Context())
+	if _, err := inv.AddNode(ctx, "anchor"); err != nil {
+		t.Fatal(err)
+	}
+	cancel() // the sender is going away: its context is cancelled between the send and the record
+	body := []byte(`{"declaration":1,"resources":[]}`)
+	digest, err := recordSent(ctx, inv, "anchor", body, nil)
+	if err != nil {
+		// NodeByName on the cancelled context may itself refuse; the record must still be possible
+		// through the detached context, so look the node up again on a live one.
+		t.Fatalf("recording a send after cancellation failed: %v", err)
+	}
+	outstanding, err := inv.Outstanding(t.Context(), "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if outstanding != digest || digest != digestOf(body) {
+		t.Fatalf("the send was not recorded: outstanding %q, sent %q", outstanding, digest)
+	}
+}

cmd/mesh-controller/licence.go

@@ -0,0 +1,481 @@
+package main
+
+import (
+	"bufio"
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"io"
+	"os"
+	"strings"
+)
+
+// licenceCommand is everything about model access the mesh holds.
+//
+// **A licence is a named thing and the name is the operator's** (novox/hq ADR 0024). *The personal
+// account*, *the organisation's account* — those are names a person uses, and the mesh has to use
+// them too, because the whole point is saying which one a given consumer uses.
+func licenceCommand(ctx context.Context, args []string) error {
+	if len(args) == 0 {
+		return errors.New(
+			"licence add|list|use|release|key|manager|set-grant|refresh|submit-refresh|forget")
+	}
+	switch args[0] {
+	case "add":
+		return licenceAdd(ctx, args[1:])
+	case "list":
+		return licenceList(ctx)
+	case "use":
+		return licenceUse(ctx, args[1:], true)
+	case "release":
+		return licenceUse(ctx, args[1:], false)
+	case "key":
+		return licenceKey(ctx, args[1:])
+	case "manager":
+		return licenceManager(ctx, args[1:])
+	case "set-grant":
+		return licenceSetGrant(ctx, args[1:])
+	case "refresh":
+		return licenceRefresh(ctx, args[1:])
+	case "submit-refresh":
+		return licenceSubmitRefresh(ctx, args[1:])
+	case "forget":
+		return licenceForget(ctx, args[1:])
+	}
+	return fmt.Errorf(
+		"licence %q; it is add, list, use, release, key, manager, set-grant, refresh, "+
+			"submit-refresh or forget", args[0])
+}
+
+func licenceAdd(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("licence add", flag.ContinueOnError)
+	// What a consumer must know that is not secret — a base URL, a model name. Never the key.
+	serves := set.String("serves", "",
+		"JSON a consumer must know that is not secret, such as a base URL or a model")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 2 {
+		return errors.New(`licence add <vendor> <name> [--serves '{"model":"..."}']`)
+	}
+	vendor, name := positionals[0], positionals[1]
+
+	values := map[string]any{}
+	if strings.TrimSpace(*serves) != "" {
+		if err := json.Unmarshal([]byte(*serves), &values); err != nil {
+			return fmt.Errorf("--serves is not JSON: %w", err)
+		}
+	}
+
+	held, err := openLicences(ctx)
+	if err != nil {
+		return err
+	}
+	defer held.Close()
+	if err := held.Add(ctx, name, vendor, values); err != nil {
+		return err
+	}
+	fmt.Printf("%s (%s) recorded. Nothing uses it yet, and it has no key:\n"+
+		"  licence use %s <node> <module>\n  licence key %s\n", name, vendor, name, name)
+	return nil
+}
+
+func licenceList(ctx context.Context) error {
+	held, err := openLicences(ctx)
+	if err != nil {
+		return err
+	}
+	defer held.Close()
+
+	all, err := held.All(ctx)
+	if err != nil {
+		return err
+	}
+	if len(all) == 0 {
+		// Said, not printed as nothing: an empty list and a failed read must never look the same.
+		fmt.Println("this mesh holds no licences")
+		return nil
+	}
+	for _, one := range all {
+		holders, err := held.HoldersOf(ctx, one.Name)
+		if err != nil {
+			return err
+		}
+		fmt.Printf("%s  (%s)\n", one.Name, one.Vendor)
+		if len(holders) == 0 {
+			fmt.Printf("  nobody uses it\n")
+		}
+		for _, h := range holders {
+			// Whether it has a key is the question somebody is actually asking, so it is said
+			// per holder rather than per licence: the key was sealed to the holders that existed
+			// when it was supplied, and one recorded afterwards has none.
+			state := "has no key — supply it again with `licence key " + one.Name + "`"
+			if h.Sealed != "" {
+				state = "has a key"
+			}
+			fmt.Printf("  %s on %s: %s\n", h.Module, h.Node, state)
+		}
+	}
+	return nil
+}
+
+func licenceUse(ctx context.Context, args []string, using bool) error {
+	verb := "use"
+	if !using {
+		verb = "release"
+	}
+	if len(args) != 3 {
+		return fmt.Errorf("licence %s <name> <node> <module>", verb)
+	}
+	name, node, module := args[0], args[1], args[2]
+
+	held, err := openLicences(ctx)
+	if err != nil {
+		return err
+	}
+	defer held.Close()
+
+	if !using {
+		if err := held.StopUsing(ctx, name, node, module); err != nil {
+			return err
+		}
+		fmt.Printf("%s on %s no longer uses %s. Its copy of the key goes on the next push\n",
+			module, node, name)
+		return nil
+	}
+	if err := held.Use(ctx, name, node, module); err != nil {
+		return err
+	}
+	fmt.Printf("%s on %s uses %s.\n", module, node, name)
+	// The consequence, said now rather than discovered as a machine that resolves and receives
+	// nothing: the mesh discarded the plaintext, so a holder added after the key was supplied has
+	// no key and the mesh cannot make one.
+	sealed, err := held.KeyFor(ctx, name, node, module)
+	if err != nil {
+		return err
+	}
+	if sealed == "" {
+		fmt.Printf("  It has no key yet — the mesh discarded the plaintext when it was supplied "+
+			"and cannot seal another. Supply it again:\n  licence key %s\n", name)
+	}
+	return nil
+}
+
+// licenceKey is the *accept* verb novox/hq ADR 0024 names as missing.
+//
+// Take a value, seal it to each holder, and discard the plaintext. Every other credential the
+// mesh handles it generated itself; an API key arrives from a person, and a mesh that kept
+// operator-supplied keys readably is the arrangement this project measured and rejected.
+func licenceKey(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("licence key", flag.ContinueOnError)
+	// A file rather than an argument, by default. A key on a command line is a key in shell
+	// history and in every process listing taken while it ran.
+	from := set.String("file", "", "read the key from a file instead of standard input")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 1 {
+		return errors.New("licence key <name> [--file <path>]")
+	}
+	name := positionals[0]
+
+	var value string
+	if *from != "" {
+		raw, err := os.ReadFile(*from)
+		if err != nil {
+			return err
+		}
+		value = strings.TrimSpace(string(raw))
+	} else {
+		fmt.Fprintln(os.Stderr, "reading the key from standard input; it is not echoed anywhere")
+		reader := bufio.NewReader(os.Stdin)
+		line, err := reader.ReadString('\n')
+		if err != nil && line == "" {
+			return fmt.Errorf("nothing was given on standard input: %w", err)
+		}
+		value = strings.TrimSpace(line)
+	}
+
+	held, err := openLicences(ctx)
+	if err != nil {
+		return err
+	}
+	defer held.Close()
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	sealed, err := held.Accept(ctx, name, value, func(node string) (string, error) {
+		return inv.SealingKeyOf(ctx, node)
+	})
+	if err != nil {
+		if sealed > 0 {
+			// Some holders got it and some did not, and the person holding the key is the only
+			// one who can finish the job. Saying how far it got is the difference between running
+			// this again knowing what it will do and running it hoping.
+			return fmt.Errorf(
+				"%w\n\n%d holder(s) were sealed before this. Running `licence key %s` again "+
+					"with the same key seals the rest and changes nothing for those already done",
+				err, sealed, name)
+		}
+		return err
+	}
+	// Not echoed back, ever. What is stored is unreadable by whoever holds it, the mesh included,
+	// and printing the value here would put the one copy that matters on a terminal.
+	fmt.Printf("sealed to %d holder(s). The mesh has discarded the key and cannot read it back\n",
+		sealed)
+	fmt.Printf("  run `push` to deliver it\n")
+	return nil
+}
+
+// licenceManager names the one node that holds a refreshable-grant licence's refresh token readably
+// and refreshes it centrally (novox/hq ADR 0050).
+//
+// **Only a refreshable-grant licence has one.** A static-key licence has no refresh token to hold, so
+// naming a manager for it is refused where the mistake is made rather than kept as a field that means
+// nothing — the absent manager is part of what keeps a static key from ever holding a value readably
+// at rest.
+func licenceManager(ctx context.Context, args []string) error {
+	if len(args) != 3 {
+		return errors.New("licence manager <name> <node> <module>")
+	}
+	name, node, module := args[0], args[1], args[2]
+
+	held, err := openLicences(ctx)
+	if err != nil {
+		return err
+	}
+	defer held.Close()
+	if err := held.SetManager(ctx, name, node, module); err != nil {
+		return err
+	}
+	fmt.Printf("%s on %s holds and refreshes %s.\n"+
+		"  Its refresh token is sealed to %s's key — readable by that node alone, not by any other "+
+		"node and not by this database. Put %s on the licence too so it is delivered the token:\n"+
+		"  licence use %s %s %s\n", module, node, name, node, module, name, node, module)
+	return nil
+}
+
+// sealedGrantJSON is the wire shape of a sealed refresh token on this command surface: an anonymous
+// sealed box and the public key it was sealed to, and nothing else.
+//
+// **Every field of it is ciphertext or a public key.** `sealed` is the refresh token as a
+// `crypto_box_seal` to the manager node's public key; `manager_key` is that public key. Neither is
+// the refresh token in the clear — which is why this surface may read one in (`set-grant`,
+// `submit-refresh`) without the control plane ever holding a refresh token it could read. The manager
+// module, on the manager node, seals it; the HOST, on that node, unseals it to deliver cleartext. This
+// database, and this surface, only ever forward the box (novox/hq ADR 0050).
+type sealedGrantJSON struct {
+	Sealed     string `json:"sealed"`
+	ManagerKey string `json:"manager_key"`
+}
+
+// readSealedGrant reads a sealed refresh token from a file or standard input as JSON.
+func readSealedGrant(from string) (sealedGrantJSON, error) {
+	var raw []byte
+	var err error
+	if from != "" {
+		raw, err = os.ReadFile(from)
+	} else {
+		raw, err = readAllStdin()
+	}
+	if err != nil {
+		return sealedGrantJSON{}, err
+	}
+	var g sealedGrantJSON
+	if err := json.Unmarshal(raw, &g); err != nil {
+		return sealedGrantJSON{}, fmt.Errorf("the sealed refresh token is not JSON: %w", err)
+	}
+	if g.Sealed == "" || g.ManagerKey == "" {
+		return sealedGrantJSON{}, errors.New(
+			"a sealed refresh token is {sealed, manager_key}, and one part is missing")
+	}
+	return g, nil
+}
+
+func readAllStdin() ([]byte, error) {
+	reader := bufio.NewReader(os.Stdin)
+	return io.ReadAll(reader)
+}
+
+// licenceSetGrant stores a sealed refresh token the manager module produced — adoption, and the
+// re-seal after a rotation done outside this process (novox/hq ADR 0050).
+//
+// **It takes a sealed box, never a refresh token.** The manager module, on the manager node, reads
+// the operator's refresh token, seals it to that node's own public key, and hands the box here. So the
+// one moment a refresh token is in the clear is on the manager node, never in the control plane — the
+// same bound the whole carve-out keeps. This surface refuses anything that is not a complete sealed
+// grant rather than storing half of one.
+func licenceSetGrant(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("licence set-grant", flag.ContinueOnError)
+	from := set.String("file", "", "read the sealed refresh token from a file instead of standard input")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 1 {
+		return errors.New("licence set-grant <name> [--file <path>]")
+	}
+	name := positionals[0]
+
+	grant, err := readSealedGrant(*from)
+	if err != nil {
+		return err
+	}
+
+	held, err := openLicences(ctx)
+	if err != nil {
+		return err
+	}
+	defer held.Close()
+	if err := held.SetRefreshGrant(ctx, name, grant.Sealed, grant.ManagerKey); err != nil {
+		return err
+	}
+	fmt.Printf("%s now holds a refresh token for %s, sealed to that node's key and readable by it "+
+		"alone.\n  the control plane stored the box without opening it; run `push` to deliver it\n",
+		"the manager", name)
+	return nil
+}
+
+// licenceSubmitRefresh publishes a refresh a MANAGER NODE already performed: the new access token is
+// sealed to every holder, and a rotated refresh token replaces the stored envelope (novox/hq ADR
+// 0050, Phase C).
+//
+// **This is the boundary the invariant rests on.** The manager runtime, on the manager node, opened
+// the at-rest envelope with that node's key, called the vendor's OAuth endpoint, and produced this:
+// the new access token in the clear, and — only if the vendor rotated it — the refresh token already
+// re-sealed at rest. This reads exactly those two things and no refresh token in the clear ever
+// reaches it, because it is never given one. The access token is sealed per holder and discarded,
+// as any accepted key is; the rotated envelope is stored opaque.
+func licenceSubmitRefresh(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("licence submit-refresh", flag.ContinueOnError)
+	accessFrom := set.String("access-file", "",
+		"read the new access token from a file instead of standard input")
+	grantFrom := set.String("grant-file", "",
+		"the rotated sealed refresh token, if the vendor rotated it; omit if it did not")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 1 {
+		return errors.New(
+			"licence submit-refresh <name> [--access-file <path>] [--grant-file <path>]")
+	}
+	name := positionals[0]
+
+	var accessToken string
+	if *accessFrom != "" {
+		raw, err := os.ReadFile(*accessFrom)
+		if err != nil {
+			return err
+		}
+		accessToken = strings.TrimSpace(string(raw))
+	} else {
+		raw, err := readAllStdin()
+		if err != nil {
+			return err
+		}
+		accessToken = strings.TrimSpace(string(raw))
+	}
+	if accessToken == "" {
+		return errors.New("no access token was given, so there is nothing to seal")
+	}
+
+	// The rotated sealed token is optional: absent, the stored refresh token is left exactly as it was.
+	var newSealed, newManagerKey string
+	if *grantFrom != "" {
+		grant, err := readSealedGrant(*grantFrom)
+		if err != nil {
+			return err
+		}
+		newSealed, newManagerKey = grant.Sealed, grant.ManagerKey
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	held, err := open.Licences(ctx)
+	if err != nil {
+		return err
+	}
+	inv := open.inventory
+
+	sealed, err := held.SubmitRefresh(ctx, name, accessToken, newSealed, newManagerKey,
+		func(node string) (string, error) {
+			return inv.SealingKeyOf(ctx, node)
+		})
+	if err != nil {
+		return err
+	}
+	rotatedNote := "the refresh token was left with its manager unchanged"
+	if newSealed != "" {
+		rotatedNote = "the rotated refresh token replaced the stored box, still readable by the " +
+			"manager node alone"
+	}
+	fmt.Printf("submitted a refresh for %s: a new access token sealed to %d holder(s), and %s.\n"+
+		"  run `push` to deliver it\n", name, sealed, rotatedNote)
+	return nil
+}
+
+// licenceRefresh mints a new access token for a refreshable-grant licence and seals it to every
+// holder (novox/hq ADR 0050). The refresh token stays with the manager and is never delivered.
+//
+// The vendor's actual refresh is a plug-in this build does not ship (Phase C), so here this reports
+// that plainly rather than pretending to have refreshed.
+func licenceRefresh(ctx context.Context, args []string) error {
+	if len(args) != 1 {
+		return errors.New("licence refresh <name>")
+	}
+	name := args[0]
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	held, err := open.Licences(ctx)
+	if err != nil {
+		return err
+	}
+	inv := open.inventory
+
+	sealed, err := held.Refresh(ctx, name, func(node string) (string, error) {
+		return inv.SealingKeyOf(ctx, node)
+	})
+	if err != nil {
+		return err
+	}
+	fmt.Printf("refreshed %s: a new access token sealed to %d holder(s), and the refresh token left "+
+		"with its manager.\n  run `push` to deliver it\n", name, sealed)
+	return nil
+}
+
+func licenceForget(ctx context.Context, args []string) error {
+	if len(args) != 1 {
+		return errors.New("licence forget <name>")
+	}
+	held, err := openLicences(ctx)
+	if err != nil {
+		return err
+	}
+	defer held.Close()
+	if err := held.Forget(ctx, args[0]); err != nil {
+		return err
+	}
+	// Said plainly, because the mesh cannot do it and pretending otherwise is worse than useless:
+	// a licence outliving its holder is a live credential nobody is watching.
+	fmt.Printf("%s is forgotten, and every record of who held it with it.\n"+
+		"  The key itself is not the mesh's to revoke — do that where the licence was bought\n",
+		args[0])
+	return nil
+}

cmd/mesh-controller/main.go

@@ -0,0 +1,321 @@
+// Command mesh-controller is the control plane: everything that needs to know about more than one
+// node (novox/hq ADR 0006).
+//
+// It runs as one process holding several contexts, each owning its own store. Today it holds one,
+// `inventory`, and does one thing with it — brings its schema up to date, which is step 3 of the
+// bootstrap in novox/hq 07-the-foundation and the step the first node cannot get past without.
+package main
+
+import (
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"os"
+	"os/signal"
+	"syscall"
+
+	"github.com/novox/mesh-controller/internal/identity"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/licences"
+	"github.com/novox/mesh-controller/internal/link"
+	"github.com/novox/mesh-controller/internal/store"
+)
+
+// version is stamped at link time. Unset in a development build, and it says so rather than
+// claiming a number.
+var version = "development build"
+
+// held is a context this process was granted, and the schema it carries.
+//
+// novox/hq ADR 0006 names seven. One is built. The list is short because the others do not exist
+// yet, not because they are optional.
+var held = []struct {
+	name       string
+	migrations func() ([]store.Migration, error)
+}{
+	{inventory.Name, inventory.Migrations},
+	{identity.Name, identity.Migrations},
+	{licences.Name, licences.Migrations},
+}
+
+func main() {
+	if err := run(); err != nil {
+		fmt.Fprintf(os.Stderr, "mesh-controller: %v\n", err)
+		os.Exit(1)
+	}
+}
+
+func run() error {
+	args := os.Args[1:]
+	if len(args) == 0 {
+		usage()
+		return fmt.Errorf("no command given")
+	}
+
+	ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM)
+	defer stop()
+
+	switch args[0] {
+	case "build":
+		return buildCommand(ctx, args[1:])
+	case "builder":
+		return builderCommand(ctx, args[1:])
+	case "board":
+		return boardCommand(ctx, args[1:])
+	case "api":
+		return apiCommand(ctx, args[1:])
+	case "licence":
+		return licenceCommand(ctx, args[1:])
+	case "rotate":
+		return rotateCommand(ctx, args[1:])
+	case "ask":
+		return askCommand(ctx, args[1:])
+	case "builds":
+		return buildsCommand(ctx, args[1:])
+	// The build queue, controlled by hand (novox/hq ADR 0219).
+	case "queue":
+		return queueCommand(ctx, args[1:])
+	case "cancel":
+		return cancelCommand(ctx, args[1:])
+	case "clear":
+		return clearCommand(ctx, args[1:])
+	case "rebuild":
+		return rebuildCommand(ctx, args[1:])
+	case "replay":
+		return replayCommand(ctx, args[1:])
+	case "kill":
+		return killCommand(ctx, args[1:])
+	case "pause", "resume":
+		return pauseCommand(ctx, args[0], args[1:])
+	case "collection":
+		return collectionCommand(ctx, args[1:])
+	case "plans":
+		return plansCommand(ctx, args[1:])
+	case "pin":
+		return pinCommand(ctx, args[1:], true)
+	case "unpin":
+		return pinCommand(ctx, args[1:], false)
+	// `prepare` is how the mesh asks any module to bring its state to the shape this version needs
+	// (novox/hq ADR 0135), and the control plane answers it the same way as everything else — its
+	// own schema is not a special case. `migrate` remains the word a person types.
+	case "prepare", "migrate":
+		return migrate(ctx)
+	case "node":
+		return nodeCommand(ctx, args[1:])
+	case "token":
+		return tokenCommand(ctx, args[1:])
+	case "identity":
+		return identityCommand(ctx, args[1:])
+	case "broker":
+		return brokerCommand(ctx, args[1:])
+	case "serve":
+		return serve(ctx)
+	case "upgrade":
+		return upgradeCommand(ctx, args[1:])
+	case "declare":
+		return declare(ctx, args[1:])
+	case "overlay":
+		return overlayCommand(ctx, args[1:])
+	case "module":
+		return moduleCommand(ctx, args[1:])
+	case "assign", "unassign":
+		return assignCommand(ctx, args[0], args[1:])
+	case "take":
+		return takeCommand(ctx, args[1:])
+	case "converge":
+		return convergeCommand(ctx, args[1:])
+	case "adopt":
+		return adoptCommand(ctx, args[1:])
+	case "settings":
+		return settingsCommand(ctx, args[1:])
+	case "secret":
+		return secretCommand(ctx, args[1:])
+	case "operator":
+		return operatorCommand(ctx, args[1:])
+	case "plan":
+		return planCommand(ctx, args[1:])
+	case "push":
+		return pushCommand(ctx, args[1:])
+	case "rollout":
+		return rolloutCommand(ctx, args[1:])
+	case "seats":
+		return seatsCommand(ctx, args[1:])
+	case "seat":
+		return seatCommand(ctx, args[1:])
+	case "status":
+		return statusCommand(ctx, args[1:])
+	case "version":
+		fmt.Println(version)
+		return nil
+	case "help", "-h", "--help":
+		usage()
+		return nil
+	default:
+		usage()
+		return fmt.Errorf("%q is not a command", args[0])
+	}
+}
+
+func usage() {
+	fmt.Fprint(os.Stderr, `mesh-controller — the control plane
+
+  migrate                        bring each context's schema up to date
+  prepare                        the same, asked the way the mesh asks any module (ADR 0135)
+  node add <name> [--adopted]    create a node record; --adopted: the machine is in use
+  node list                      the nodes this mesh knows about
+  node show <name>               what one machine reported it can do, and why
+  node public-domain <name>      the domain it composes its routed names under
+  node public-domain <name> <d>  ...set it to d
+  node public-domain <name> --clear   ...it faces the outside no longer
+  node networks <name>           the networks it routes for what it hosts
+  node networks <name> <cidr>... ...set them; its filter forwards these too
+  node networks <name> --clear   ...only the container runtime's own
+  token issue --node <name>      a one-time right to join, for an existing record
+  token issue --new <name>       create the record and issue for it
+  token issue ... --adopted      ...for a machine in use, which joins adopted
+  identity show                  this control plane's signing key
+  broker show                    where the broker is, and what to expect there
+  serve                          consume what nodes say, and answer
+  declare <node> <file>          send a node a signed declaration
+  overlay place <node> [flags]   say where a node is and how it is reached
+  overlay show                   the private network, as the mesh computes it
+  module add <file>              register a module from its manifest
+  module check <file|dir>...     judge manifests where they are written, with no mesh (exit 1 on any problem)
+  module list                    what modules this mesh knows about
+  module moved <name> <commit>   the source has a newer commit than the mesh built
+  module forget <name>           remove one, unless a node runs it or the mesh holds things for it
+  module forget <name> --and-what-it-holds   ...and discard its settings, secrets and ports too
+  module issue <name> --node <m> a broker account for a module, scoped to its emits and consumes
+  upgrade <name>                 what happens when this module's current version moves
+  upgrade <name> roll-out [--together]   ...send it to the machines running it
+  upgrade <name> record          ...record that they are behind, and send nothing
+  status [--json]                what is wrong, what is quiet, and what is out of date
+  seats [--json]                 every seat this mesh defines, what it delivers, and who holds it
+  seat rename <from> <to>        rename a seat; its former name still resolves (ADR 0122)
+  seat <name> --to <node>/<module>   hand a seat to that assignment as one act; never empty in between (ADR 0131)
+  seat <name> --add <node>/<module>  add a holder beside the others, for a replicated seat (ADR 0223)
+  board [--listen ADDR]          the same three questions, as a page that holds nothing
+  api --issuer URL [--listen A]  assign and unassign over http, for a surface that is not here
+  assign <node> <module>...      put modules on a node, judged together (ADR 0207)
+  unassign <node> <module>...    take them off
+  take <node> <module>           preview a module's cutover on an adopted node: what runs beside
+                                 what it declares; --yes <digest> cuts it over as previewed
+  converge <node> [--yes <digest>] [--filter nftables]   preview, then make, an adopted node converged
+  adopt <node>                   return a converged node to adopted; what was taken stays taken
+  settings set <module> <file>   what a module's config should say, for the whole mesh
+  settings set <module> <file> --node <n>   ...or for one machine
+  settings clear <module> [--node <n>]      take a layer away
+  secret accept <node> <module> <name>     carry a value the mesh did not make and cannot invent
+  secret accept ... --from <file>           ...read it from a file rather than being asked
+  secret recover <node> <module> <name> --key <operator-key> [--out <file>] [--from-export <file>]
+                                            break-glass: open the operator-sealed copy, to a 0600 file
+  secret export [--out <file>]              every operator-sealed copy, ciphertext — keep it with the key
+  operator key make [--out <file>]          make the operator's sealing key, off the mesh; private half to the file only
+  operator key set <public> [--replace]     tell the mesh which operator key to seal to
+  operator key show                         the operator key, and what it can recover
+  build <repository> [--ref R]   have a build machine build it, and record what came out
+  build --behind                 build every module the mesh holds older than its source
+  build --on <module>            rebuild every module that stands on this module's artifacts, bases first
+  builds [<module>]              what has been built lately, and what came of it
+  queue [--json]                 every ask in the build queue: waiting, in flight (where, how long), dead
+  cancel <id>                    drop a waiting or dead ask; recorded failed, cancelled by hand
+  clear [--dead]                 cancel every waiting ask (and the dead ones); never one in flight
+  rebuild <module|build-id>      ask the module's source again, or that build's, under a new id
+  replay <build-id> [--register [--older]]   that build's commit again; a dry run unless --register
+  kill <id>                      end a build where it runs; recorded failed, killed by hand
+  pause [<node>] / resume [<node>]   the build seat's holder there, or every holder, takes nothing new / again
+  plans retry <id>               ask a failed plan's failed builds again, and carry the plan on
+  collection [--json]            kept archives held/unheld by a manifest, and what the sweep may let go
+  builder issue <name>           a broker account for a build machine, scoped to build work,
+                                 delivered as the builder module's broker secret (module add it first)
+  licence add|list|use|key       model access, under the name a person calls it
+  licence manager <name> <node>  the node that holds a refreshable licence's refresh token
+  licence refresh <name>         mint a new access token and seal it to every holder
+  rotate <provision> [--consumer <n>]  a new credential for every holder, both ends at once
+  ask <module> <tool> [json]           call one of a module's tools over the broker, and print its answer
+  pin <node> <provision> <from-node> <module>
+                                 which provider this one gets a provision from: the module, and its node
+  unpin <node> <provision>       put that question back
+  plan <node> [--files|--json]     what that node would run, and why
+  push [<node>] [--behind]       send a node everything it should be, or only those that need it
+  version                        what this binary is
+
+Each context reaches its own store through its own credential (novox/hq ADR 0008), named
+`+store.Variable("<context>")+` — or `+store.FileVariable("<context>")+`, naming a file that holds
+the same thing and keeps the password out of the environment. This process holds:
+
+`)
+	for _, c := range held {
+		fmt.Fprintf(os.Stderr, "  %-12s database %-12s from %s\n",
+			c.name, store.Database(c.name), store.Variable(c.name))
+	}
+	fmt.Fprintln(os.Stderr)
+}
+
+// parseAround reads flags that may sit before, after or between positional arguments.
+//
+// The standard library stops at the first non-flag argument, so `module add thing.json --source x`
+// parses no flags at all and silently ignores every one of them. The host learned this the same
+// way and says so in its own parser: a flag that is quietly dropped is the fault this project
+// keeps naming, and it looks exactly like success.
+func parseAround(set *flag.FlagSet, args []string) ([]string, error) {
+	var positionals []string
+	rest := args
+	for {
+		if err := set.Parse(rest); err != nil {
+			return nil, err
+		}
+		rest = set.Args()
+		if len(rest) == 0 {
+			return positionals, nil
+		}
+		positionals = append(positionals, rest[0])
+		rest = rest[1:]
+	}
+}
+
+// Built is the daemon hearing a build's outcome on the bus — its own asking, an announcement's, or
+// a tool's that did not wait (novox/hq issue 176) — and taking it in: recorded, and the module
+// registered, the same as the waiting command does. Said either way, so the daemon's log tells what
+// became of a build nobody was watching.
+func (b builds) Built(ctx context.Context, result link.BuildResult) error {
+	// **A dry run is looked at, never taken in** (novox/hq issue 240). On 2026-10-04 a dry run of an
+	// unmerged branch was heard here like any build, registered, and its definition reached a machine
+	// before anyone had reviewed it.
+	if result.DryRun {
+		fmt.Printf("%s: a dry run of %s on %s, not taken in\n", result.ID, result.Repository, result.Ref)
+		return nil
+	}
+	manifest, _, err := takeIn(ctx, b.inv, result)
+	// When it was asked, so a plan takes as its outcome only a build asked for it or after it
+	// (novox/hq 04-ISSUES/219). Zero when the id does not say.
+	asked, _ := link.BuildAskedAt(result.ID)
+	switch {
+	case err != nil && result.Failed != "":
+		fmt.Printf("%s: %v\n", result.ID, err)
+		if result.Module != "" {
+			planBuilt(ctx, b.open, result.Module, result.Commit, result.Failed, asked, result.ID)
+		} else {
+			planFailedBuild(ctx, b.open, result)
+		}
+		return nil
+	case errors.Is(err, inventory.ErrSuperseded):
+		// Not a failure: the module is already at what a later request built. A plan that asked
+		// before that later request is answered by it; one that asked after it ignores this.
+		fmt.Printf("%s: %v\n", result.ID, err)
+		planBuilt(ctx, b.open, manifest.Module, result.Commit, "", asked, result.ID)
+		return nil
+	case err != nil:
+		fmt.Printf("%s: heard and recorded, and not registered: %v\n", result.ID, err)
+		if manifest.Module != "" {
+			planBuilt(ctx, b.open, manifest.Module, result.Commit, err.Error(), asked, result.ID)
+		}
+		return nil
+	}
+	fmt.Printf("%s: %s %s registered, built on %s from %s\n",
+		result.ID, manifest.Module, manifest.Version, result.On, short(result.Commit))
+	saysWhenThePolicyActs(ctx, b.inv, manifest.Module)
+	planBuilt(ctx, b.open, manifest.Module, result.Commit, "", asked, result.ID)
+	return nil
+}

cmd/mesh-controller/mesh_for_test.go

@@ -0,0 +1,108 @@
+package main
+
+import (
+	"crypto/ecdh"
+	"crypto/rand"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/licences"
+	"github.com/novox/mesh-controller/internal/overlay"
+)
+
+// A mesh a command can be run against.
+//
+// The commands here were tested through the pieces they call and never through themselves, so
+// three faults that only exist where the pieces meet — an assignment reported as fine while it
+// blocked other machines, a read-shaped invocation that wrote, a JSON interface that stopped
+// emitting JSON — were invisible to every test in this package. This raises the real stores and
+// calls the real functions.
+
+// aMesh is two placed, capable machines on a private network, with nothing assigned but the
+// network itself.
+//
+// `anchor` is the hub. That is not decoration: a mesh whose hub cannot be resolved has no private
+// network at all, which is how one machine's problem reaches every other.
+func aMesh(t *testing.T) *stores {
+	t.Helper()
+	inventory.ForTest(t) // raises the store, migrates it, and points the environment at it
+	licences.ForTest(t)  // planning reaches this one too, by name and never by connection
+
+	open, err := openStores(t.Context())
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(open.Close)
+	for _, m := range provided {
+		if err := open.inventory.Provide(t.Context(), m); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	for i, name := range []string{"anchor", "laptop"} {
+		record, err := open.inventory.AddNode(t.Context(), name)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if err := open.inventory.SetPlace(t.Context(), name, name+".example:51820", "here",
+			name == "anchor", fmt.Sprintf("10.77.0.%d", i+1)); err != nil {
+			t.Fatal(err)
+		}
+		reported, err := json.Marshal(map[string]any{"capabilities": []map[string]any{
+			{"name": "container-runtime", "present": true},
+			{"name": "wireguard", "present": true},
+			{"name": "systemd", "present": true},
+		}})
+		if err != nil {
+			t.Fatal(err)
+		}
+		var profile map[string]any
+		if err := json.Unmarshal(reported, &profile); err != nil {
+			t.Fatal(err)
+		}
+		if err := open.inventory.RecordProfile(t.Context(), record.ID, profile); err != nil {
+			t.Fatal(err)
+		}
+		if err := open.inventory.RecordSealingKey(t.Context(), record.ID, aPublicKey(t)); err != nil {
+			t.Fatal(err)
+		}
+		if err := open.inventory.RecordOverlayKey(t.Context(), record.ID, aPublicKey(t)); err != nil {
+			t.Fatal(err)
+		}
+		if _, err := open.inventory.Assign(t.Context(), name, overlay.Name); err != nil {
+			t.Fatal(err)
+		}
+	}
+	return open
+}
+
+// aPublicKey is a key a machine could have reported. Its private half is thrown away: nothing here
+// opens anything, it only needs the mesh to believe a machine has a key.
+func aPublicKey(t *testing.T) string {
+	t.Helper()
+	k, err := ecdh.X25519().GenerateKey(rand.Reader)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return base64.StdEncoding.EncodeToString(k.PublicKey().Bytes())
+}
+
+// register puts a manifest in the catalogue.
+func register(t *testing.T, open *stores, m catalogue.Manifest) {
+	t.Helper()
+	if err := open.inventory.RegisterModule(t.Context(), m, inventory.Source{}); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// rivals are two modules that cannot share a machine, which is the shortest way to make a node's
+// own set of assignments incoherent using nothing but commands a person has.
+func rivals() (catalogue.Manifest, catalogue.Manifest) {
+	claim := []catalogue.Claim{{Name: "the-seat", Scope: catalogue.ScopeNode}}
+	return catalogue.Manifest{Module: "rival-one", Version: "1", Claims: claim},
+		catalogue.Manifest{Module: "rival-two", Version: "1", Claims: claim}
+}

cmd/mesh-controller/missed_merges.go

@@ -0,0 +1,131 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// **A merge the bus announced and the controller never acted on is caught up** (novox/hq issue 266).
+//
+// The forge's poll announces every merge on the events stream, and the controller acts on what its
+// consumer there hands it. On 2026-10-06 one merge was on the stream and never handed over: the bus
+// server moved the consumer past it — a fault of consumers with several filters in the server the
+// mesh ran — and the controller, which only acts on what it is handed, said nothing. The modules
+// built from that repository stayed behind and were built by hand.
+//
+// So the stream is read back on a timer, on a consumer of its own filtered on merges alone, and every
+// announcement older than mergeGrace is judged as SourceMoved would judge it. **No record of what
+// was handled is kept, because none is needed**: acting on a merge marks every module it moved as
+// looked at since, so an announcement already acted on reads as history and moves nothing. One that
+// would still move something was never acted on — it is said, and acted on now.
+const (
+	// mergeGrace is how long an announcement is left to the controller's own consumer before it is
+	// judged missed. That consumer hands over one event at a time, and a merge waits behind a build
+	// outcome that is being acted on; acting on a merge itself asks builds and does not wait for them.
+	mergeGrace = 10 * time.Minute
+	// mergeLookBack is how far back a pass reads. A merge missed longer ago than this was missed by a
+	// controller that was not running this, and is the operator's to look at, not a surprise rebuild.
+	mergeLookBack = 24 * time.Hour
+	// mergeCatchUpEvery is how often the stream is read back.
+	mergeCatchUpEvery = 5 * time.Minute
+)
+
+// merges is what reads back the forge's announcements; the link server, or a test's list.
+type merges interface {
+	AnnouncedMerges(ctx context.Context, since time.Time) ([]link.AnnouncedMerge, error)
+}
+
+// catchingUpOnMerges reads back the forge's announcements on a timer, until the context ends.
+func catchingUpOnMerges(ctx context.Context, open *stores, announced merges) {
+	f := following{open}
+	catalogued := func(ctx context.Context) ([]inventory.Entry, map[string][]inventory.ReadRepository, error) {
+		entries, err := open.inventory.Catalogued(ctx)
+		if err != nil {
+			return nil, nil, err
+		}
+		read, err := open.inventory.ReadRepositories(ctx)
+		return entries, read, err
+	}
+	failing := ""
+	tick := time.NewTicker(mergeCatchUpEvery)
+	defer tick.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-tick.C:
+		}
+		err := catchUpOnMerges(ctx, time.Now(), announced, catalogued, f.SourceMoved, func(format string, args ...any) {
+			fmt.Printf(format+"\n", args...)
+		})
+		// A pass that cannot read says so once, not every five minutes, and says when it reads again.
+		why := ""
+		if err != nil {
+			why = err.Error()
+		}
+		if why != failing {
+			if why != "" {
+				fmt.Printf("merges the bus may not have handed over cannot be looked for: %s\n", why)
+			} else {
+				fmt.Println("merges the bus may not have handed over are looked for again")
+			}
+			failing = why
+		}
+	}
+}
+
+// catchUpOnMerges is one pass: every announcement older than mergeGrace that acting on would still
+// move something is said and acted on, oldest first.
+//
+// Judged twice: once against the catalogue as the pass found it, and again just before acting,
+// because acting on an earlier missed merge of the same repository may have moved what a later one
+// would have.
+func catchUpOnMerges(ctx context.Context, now time.Time, announced merges,
+	catalogued func(context.Context) ([]inventory.Entry, map[string][]inventory.ReadRepository, error),
+	act func(context.Context, link.SourceMoved) error, say func(string, ...any)) error {
+
+	all, err := announced.AnnouncedMerges(ctx, now.Add(-mergeLookBack))
+	if err != nil {
+		return err
+	}
+	entries, read, err := catalogued(ctx)
+	if err != nil {
+		return err
+	}
+	for _, a := range all {
+		if now.Sub(a.At) < mergeGrace {
+			continue
+		}
+		if len(wouldMove(a.SourceMoved, entries, read)) == 0 {
+			continue
+		}
+		if entries, read, err = catalogued(ctx); err != nil {
+			return err
+		}
+		moves := wouldMove(a.SourceMoved, entries, read)
+		if len(moves) == 0 {
+			continue
+		}
+		var names []string
+		for _, e := range moves {
+			names = append(names, e.Manifest.Module)
+		}
+		say("%s/%s merged into %s (%.8s), announced %s ago, and the controller never acted on it: the bus "+
+			"did not hand the announcement over (novox/hq issue 266). %s %s behind it; acting on it now",
+			a.Owner, a.Repo, a.Base, a.Commit, now.Sub(a.At).Round(time.Minute), readableList(names),
+			isAre(len(names)))
+		if err := act(ctx, a.SourceMoved); err != nil {
+			say("%s/%s moved to %.8s and the mesh could not act on it: %v; the next pass tries again",
+				a.Owner, a.Repo, a.Commit, err)
+			continue
+		}
+		if entries, read, err = catalogued(ctx); err != nil {
+			return err
+		}
+	}
+	return nil
+}

cmd/mesh-controller/missed_merges_test.go

@@ -0,0 +1,180 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// announcedList is the events stream's merges as a test gives them.
+type announcedList []link.AnnouncedMerge
+
+func (a announcedList) AnnouncedMerges(_ context.Context, since time.Time) ([]link.AnnouncedMerge, error) {
+	var out []link.AnnouncedMerge
+	for _, m := range a {
+		if !m.At.Before(since) {
+			out = append(out, m)
+		}
+	}
+	return out, nil
+}
+
+// aCatalogue is what the inventory holds, changed the way acting on a merge changes it: every module
+// the merge moved is marked as looked at (inventory.SourceMoved writes source_seen = now()).
+type aCatalogue struct {
+	entries []inventory.Entry
+	acted   []string
+	fail    error
+}
+
+func (c *aCatalogue) read(context.Context) ([]inventory.Entry, map[string][]inventory.ReadRepository, error) {
+	return append([]inventory.Entry(nil), c.entries...), nil, nil
+}
+
+func (c *aCatalogue) act(now func() time.Time) func(context.Context, link.SourceMoved) error {
+	return func(_ context.Context, m link.SourceMoved) error {
+		if c.fail != nil {
+			return c.fail
+		}
+		c.acted = append(c.acted, m.Repo+"@"+m.Commit[:8])
+		for _, moved := range wouldMove(m, c.entries, nil) {
+			for i := range c.entries {
+				if c.entries[i].Manifest.Module == moved.Manifest.Module {
+					c.entries[i].Source.Head = m.Commit
+					c.entries[i].Source.Seen = now()
+				}
+			}
+		}
+		return nil
+	}
+}
+
+func at(s string) time.Time {
+	t, err := time.Parse(time.RFC3339, s)
+	if err != nil {
+		panic(err)
+	}
+	return t
+}
+
+func announced(repo, commit, mergedAt, onTheBus string, paths ...string) link.AnnouncedMerge {
+	return link.AnnouncedMerge{
+		SourceMoved: link.SourceMoved{Owner: "novox", Repo: repo, Base: "main", Commit: commit,
+			MergedAt: mergedAt, Paths: paths,
+			CloneURL: "http://forge.internal:20000/novox/" + repo + ".git"},
+		At: at(onTheBus),
+	}
+}
+
+func built(module, repo, path, commit, seen string) inventory.Entry {
+	e := fromRepo(module, "http://forge.internal:20000/novox/"+repo+".git", path)
+	e.Source.BuiltFrom, e.Source.Head, e.Source.Seen = commit, commit, at(seen)
+	return e
+}
+
+// **novox/hq issue 266, as it happened.** The forge announced a merge of the tools repository on the
+// events stream; the bus never handed it to the controller, which acted on the merges around it and
+// not on this one, and said nothing. Read back from the stream, it is the one merge that would still
+// move something — so it is said and acted on, once, and only after the controller's own consumer
+// has had its time with it.
+func TestAMergeTheBusNeverHandedOverIsActedOnLate(t *testing.T) {
+	cat := &aCatalogue{entries: []inventory.Entry{
+		built("mesh-tools", "mesh-tools", "", "8b789578aaaaaaaa", "2026-10-04T15:24:32Z"),
+		built("node-tools", "mesh-tools", "node-tools", "8b789578aaaaaaaa", "2026-10-04T15:24:32Z"),
+		// Acted on when it was announced: looked at after it was merged.
+		built("gitea", "mesh-catalog", "modules/gitea", "5c2157b8bbbbbbbb", "2026-10-05T22:39:21Z"),
+	}}
+	stream := announcedList{
+		// Nothing the mesh holds is built from the records repository.
+		announced("hq", "88f7f79fcccccccc", "2026-10-05T22:43:00Z", "2026-10-05T22:43:04Z", "04-ISSUES/x.md"),
+		// Acted on: its module was looked at since.
+		announced("mesh-catalog", "78328d4adddddddd", "2026-10-05T22:39:00Z", "2026-10-05T22:39:21Z", "modules/gitea/x.ts"),
+		// Never handed over.
+		announced("mesh-tools", "9730bd89c3e48d0e", "2026-10-05T22:46:47Z", "2026-10-05T22:47:06Z",
+			"node-tools/internal/console/console.go"),
+	}
+	var said []string
+	say := func(format string, args ...any) { said = append(said, fmt.Sprintf(format, args...)) }
+	clock := at("2026-10-05T22:50:00Z")
+	now := func() time.Time { return clock }
+	pass := func() {
+		t.Helper()
+		if err := catchUpOnMerges(context.Background(), clock, stream, cat.read, cat.act(now), say); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	pass()
+	if len(cat.acted) != 0 {
+		t.Fatalf("a merge three minutes old was taken from the controller's own consumer: %v", cat.acted)
+	}
+
+	clock = at("2026-10-05T22:58:00Z")
+	pass()
+	if strings.Join(cat.acted, ",") != "mesh-tools@9730bd89" {
+		t.Fatalf("acted on %v, wanted the one merge never handed over", cat.acted)
+	}
+	if len(said) != 1 || !strings.Contains(said[0], "novox/mesh-tools merged into main (9730bd89)") ||
+		!strings.Contains(said[0], "mesh-tools and node-tools are behind it") {
+		t.Fatalf("the missed merge was not said as one: %q", said)
+	}
+
+	clock = at("2026-10-05T23:03:00Z")
+	pass()
+	if len(cat.acted) != 1 || len(said) != 1 {
+		t.Fatalf("a merge acted on was acted on again: %v %q", cat.acted, said)
+	}
+}
+
+// A merge that changed none of the held modules' files moves nothing, so it is never "missed"; one
+// that could not be acted on is said and tried again on the next pass.
+func TestAMissedMergeThatCouldNotBeActedOnIsTriedAgain(t *testing.T) {
+	cat := &aCatalogue{
+		entries: []inventory.Entry{built("gitea", "mesh-catalog", "modules/gitea", "5c2157b8bbbbbbbb", "2026-10-05T20:00:00Z")},
+		fail:    errors.New("the store is restarting"),
+	}
+	stream := announcedList{
+		announced("mesh-catalog", "aaaaaaaa11111111", "2026-10-05T21:00:00Z", "2026-10-05T21:00:10Z",
+			"modules/plex/module.json", "modules/plex/x.ts"),
+		announced("mesh-catalog", "bbbbbbbb22222222", "2026-10-05T21:10:00Z", "2026-10-05T21:10:10Z", "modules/gitea/x.ts"),
+	}
+	var said []string
+	say := func(format string, args ...any) { said = append(said, fmt.Sprintf(format, args...)) }
+	clock := at("2026-10-05T22:00:00Z")
+	now := func() time.Time { return clock }
+	if err := catchUpOnMerges(context.Background(), clock, stream, cat.read, cat.act(now), say); err != nil {
+		t.Fatal(err)
+	}
+	if len(said) != 2 || !strings.Contains(said[1], "could not act on it") {
+		t.Fatalf("a failed catch-up was not said: %q", said)
+	}
+	cat.fail = nil
+	clock = at("2026-10-05T22:05:00Z")
+	if err := catchUpOnMerges(context.Background(), clock, stream, cat.read, cat.act(now), say); err != nil {
+		t.Fatal(err)
+	}
+	if strings.Join(cat.acted, ",") != "mesh-catalog@bbbbbbbb" {
+		t.Fatalf("acted on %v, wanted only the merge that changed a held module", cat.acted)
+	}
+}
+
+// A merge older than the look-back is left to the operator: a controller that did not run this
+// missed it, and acting on it days later would be a surprise rebuild.
+func TestAMergeOlderThanTheLookBackIsLeftAlone(t *testing.T) {
+	cat := &aCatalogue{entries: []inventory.Entry{built("gitea", "mesh-catalog", "modules/gitea", "5c2157b8bbbbbbbb", "2026-10-01T00:00:00Z")}}
+	stream := announcedList{announced("mesh-catalog", "cccccccc33333333", "2026-10-03T00:00:00Z", "2026-10-03T00:00:05Z", "modules/gitea/x.ts")}
+	clock := at("2026-10-05T22:00:00Z")
+	if err := catchUpOnMerges(context.Background(), clock, stream, cat.read, cat.act(func() time.Time { return clock }),
+		func(string, ...any) {}); err != nil {
+		t.Fatal(err)
+	}
+	if len(cat.acted) != 0 {
+		t.Fatalf("a merge of three days ago was acted on: %v", cat.acted)
+	}
+}

cmd/mesh-controller/modules.go

@@ -0,0 +1,796 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"net"
+	"os"
+	"sort"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/overlay"
+)
+
+// the catalogue: what exists, what is assigned, and how it is configured.
+//
+// Split out of main.go, which had reached 2,769 lines because appending was always the
+// cheapest next step. That is how novox/hq ADR 0001 records `hal/sdk` reaching 34,636:
+// nothing in it was wrong, and no one edit was the one that should have been a new file.
+
+// provided is what comes with the control plane rather than from a repository.
+//
+// WireGuard, the names, and the domain module over both. The first two are here because the code
+// that works out their files is here:
+// a peer list is derived from every machine at once, so it cannot be written in a manifest, and
+// whatever computes it has to live wherever the whole picture is.
+//
+// **It is a module in every other respect** — assigned, unassigned, resolved, settled, and absent
+// from a machine nobody gave it to.
+func providedModules() []catalogue.Manifest {
+	var out []catalogue.Manifest
+	for _, raw := range []map[string]any{
+		// **Two used to be here and are gone**: one wrote the mesh's names into a hosts file, the
+		// other wrote the same machines as wildcards for a resolver to read. Neither ran software
+		// and neither could be swapped for anything, which is the test of whether a thing is a
+		// module at all (novox/hq ADR 0040). They existed because computed output needed somewhere
+		// to live, and now a module says where it wants it — `facts` in its own manifest.
+		//
+		// **And the bundle that required it is gone** (novox/hq ADR 0226): `networking` named this
+		// module's requirement and nothing else, so every machine carried two modules for one
+		// network. A machine is assigned the private network itself; what a release stops shipping
+		// is retired at the next start (stores.go, Inventory.RetireUnshipped).
+		overlay.Manifest(),
+	} {
+		var m catalogue.Manifest
+		b, _ := json.Marshal(raw)
+		_ = json.Unmarshal(b, &m)
+		out = append(out, m)
+	}
+	return out
+}
+
+var provided = providedModules()
+
+func moduleCommand(ctx context.Context, args []string) error {
+	if len(args) == 0 {
+		return errors.New("module add <file>, module check <file>..., module list, or module forget <name>")
+	}
+	// `check` needs no mesh, and must not: it is what somebody runs in their own repository before
+	// there is a mesh in reach (novox/hq issue 148). A directory expands to every manifest under it.
+	if args[0] == "check" {
+		var paths []string
+		for _, a := range args[1:] {
+			if info, err := os.Stat(a); err == nil && info.IsDir() {
+				under, err := manifestsUnder(a)
+				if err != nil {
+					return err
+				}
+				paths = append(paths, under...)
+				continue
+			}
+			paths = append(paths, a)
+		}
+		return moduleCheck(paths, os.Stdout)
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	switch args[0] {
+	case "add":
+		set := flag.NewFlagSet("module add", flag.ContinueOnError)
+		repo := set.String("source", "", "where this module comes from")
+		ref := set.String("ref", "", "the branch followed there")
+		commit := set.String("commit", "", "the commit this manifest was read at")
+		// **Where inside the repository the module is** (novox/hq ADR 0069). A module is a
+		// repository *and* a directory, and a record that carries only the repository names a
+		// module.json at its root — so every later build of it looks in the wrong place and fails
+		// with "no module.json at its root". Nine modules on this mesh were registered that way
+		// and none of them could be rebuilt (2026-09-28).
+		path := set.String("path", "", "the module's directory inside that repository")
+		self := set.Bool("self", false, "the source is a path on the forge holding the git seat")
+		positionals, err := parseAround(set, args[1:])
+		if err != nil {
+			return err
+		}
+		if len(positionals) != 1 {
+			return errors.New("module add <manifest.json> [--source <repo> [--self] [--path P] " +
+				"--ref <branch> --commit <sha>]")
+		}
+		raw, err := os.ReadFile(positionals[0])
+		if err != nil {
+			return err
+		}
+		m, err := catalogue.ParseManifest(raw)
+		if err != nil {
+			return err
+		}
+		from, err := whereItComesFrom(*repo, *ref, *commit, *path, *self)
+		if err != nil {
+			return err
+		}
+		if err := namesNoInstallation(m); err != nil {
+			return err
+		}
+		if err := inv.RegisterModule(ctx, m, from); err != nil {
+			return err
+		}
+		fmt.Printf("%s registered", m.Module)
+		if *commit != "" {
+			fmt.Printf(" from %s", short(*commit))
+		}
+		if *repo != "" && *path == "" {
+			// Said, not refused: a module really at the root is the ordinary case for a repository
+			// of its own. But a repository holding many modules and a record naming none of them is
+			// a module nothing can rebuild, and the person adding it is the one who knows which.
+			fmt.Printf("\n  no directory inside %s, so it is built from that repository's root — "+
+				"`--path` if the module lives in a directory there", *repo)
+		}
+		if len(m.Provides) > 0 {
+			fmt.Printf(", providing %s", describeOffers(m.Provides))
+		}
+		fmt.Println()
+		for _, c := range m.Claims {
+			fmt.Printf("  claims %s, one per %s\n", c.Name, c.At())
+		}
+		return nil
+
+	case "list":
+		// The catalogue: what exists, where it came from, whether it is current, and who runs it.
+		// The provenance was recorded from the first build and nothing showed it, which made
+		// "is this current?" a question you could only answer by reading the database.
+		entries, err := inv.Catalogued(ctx)
+		if err != nil {
+			return err
+		}
+		// **The same list, for something other than a person** (novox/hq ADR 0195): what each module
+		// is, where it runs, whether it is current, and what it says of itself.
+		if len(args) > 1 && args[1] == "--json" {
+			type listed struct {
+				Module   string `json:"module"`
+				Version  string `json:"version"`
+				Built    string `json:"built,omitempty"`
+				Head     string `json:"head,omitempty"`
+				Current  bool   `json:"current"`
+				Provided bool   `json:"provided,omitempty"`
+				// Tools says whether the module answers tools anywhere it runs: a list of its own,
+				// a bundle the runtime serves, or a seat's verbs it claims (novox/hq ADR 0197) —
+				// what the console checks the bus's answers against.
+				Tools        bool     `json:"tools"`
+				On           []string `json:"on"`
+				Provides     []string `json:"provides,omitempty"`
+				Requires     []string `json:"requires,omitempty"`
+				Claims       []string `json:"claims,omitempty"`
+				Capabilities []string `json:"capabilities,omitempty"`
+			}
+			out := make([]listed, 0, len(entries))
+			for _, e := range entries {
+				m := e.Manifest
+				l := listed{Module: m.Module, Version: m.Version, Built: e.Source.BuiltFrom, Head: e.Source.Head,
+					Current: e.Provided || e.Source.Repository == "" || e.Source.Current(), Provided: e.Provided,
+					On: append([]string{}, e.On...), Provides: m.Offers(), Requires: m.Requires,
+					Capabilities: m.Capabilities, Tools: declaresTools(m)}
+				for _, c := range m.Claims {
+					l.Claims = append(l.Claims, c.At()+"/"+c.Name)
+				}
+				out = append(out, l)
+			}
+			return printJSON(out)
+		}
+		if len(entries) == 0 {
+			fmt.Println("this mesh knows about no modules yet")
+			return nil
+		}
+		var stale int
+		for _, e := range entries {
+			m := e.Manifest
+			fmt.Printf("%-18s %-8s", m.Module, m.Version)
+
+			switch {
+			case e.Provided:
+				fmt.Printf(" %-22s", "with the control plane")
+			case e.Source.Repository == "":
+				// Handed over by hand. Legitimate — it is how a module is fixed in a hurry — and
+				// worth saying, because nothing can rebuild it.
+				fmt.Printf(" %-22s", "handed over")
+			case !e.Source.Current():
+				stale++
+				fmt.Printf(" %-22s", "behind "+short(e.Source.BuiltFrom)+" < "+short(e.Source.Head))
+			default:
+				fmt.Printf(" %-22s", "built "+short(e.Source.BuiltFrom))
+			}
+
+			if len(e.On) > 0 {
+				fmt.Printf("  on %s", strings.Join(e.On, ", "))
+			} else {
+				fmt.Printf("  on nothing")
+			}
+			fmt.Println()
+
+			var says []string
+			if len(m.Provides) > 0 {
+				says = append(says, "provides "+describeOffers(m.Provides))
+			}
+			if len(m.Requires) > 0 {
+				says = append(says, "requires "+strings.Join(m.Requires, ", "))
+			}
+			for _, c := range m.Claims {
+				says = append(says, "claims "+c.At()+"/"+c.Name)
+			}
+			if len(m.Capabilities) > 0 {
+				says = append(says, "needs "+strings.Join(m.Capabilities, ", "))
+			}
+			if len(says) > 0 {
+				fmt.Printf("    %s\n", strings.Join(says, "  ·  "))
+			}
+		}
+		if stale > 0 {
+			fmt.Printf("\n%d module(s) behind their source — `build --behind` to catch up\n", stale)
+		}
+		return nil
+
+	case "moved":
+		if len(args) != 3 {
+			return errors.New("module moved <name> <commit> — the source has a newer commit")
+		}
+		if err := inv.SourceMoved(ctx, args[1], args[2]); err != nil {
+			return err
+		}
+		from, err := inv.SourceOf(ctx, args[1])
+		if err != nil {
+			return err
+		}
+		if from.Current() {
+			fmt.Printf("%s is current at %s\n", args[1], short(from.Head))
+			return nil
+		}
+		fmt.Printf("%s is behind: the mesh holds %s and the source has %s\n",
+			args[1], short(from.BuiltFrom), short(from.Head))
+		fmt.Printf("  run `build %s` to catch up\n", from.Repository)
+		return nil
+
+	case "forget":
+		// **What goes with it is said before it goes** (novox/hq 04-ISSUES/017). The settings, the
+		// module's own secrets and the ports the mesh chose all cascade off the module row, so
+		// `forget` used to destroy them and report "forgotten" — an action succeeding into a state
+		// its own verify would reject, and a sealed secret is not recoverable afterwards.
+		set := flag.NewFlagSet("module forget", flag.ContinueOnError)
+		andHeld := set.Bool("and-what-it-holds", false,
+			"discard its settings, its own secrets and its ports along with it")
+		positionals, err := parseAround(set, args[1:])
+		if err != nil {
+			return err
+		}
+		if len(positionals) != 1 {
+			return errors.New("module forget <name> [--and-what-it-holds]")
+		}
+		if !*andHeld {
+			if err := inv.ForgetModule(ctx, positionals[0]); err != nil {
+				return err
+			}
+			fmt.Printf("%s forgotten\n", positionals[0])
+			return nil
+		}
+		held, err := inv.DiscardModule(ctx, positionals[0])
+		if err != nil {
+			return err
+		}
+		fmt.Printf("%s forgotten\n", positionals[0])
+		for _, line := range held.Lines() {
+			// Said after the fact as well as before it: this is the only record that these
+			// existed, and the next person to ask why the module came back empty reads it here.
+			fmt.Printf("  discarded%s\n", strings.TrimPrefix(line, " "))
+		}
+		return nil
+
+	case "issue":
+		// A module's broker account, scoped by its emits and consumes (novox/hq ADR 0043) and
+		// sealed to the machine that will run it — the generic case the builder was the first of.
+		set := flag.NewFlagSet("module issue", flag.ContinueOnError)
+		forNode := set.String("node", "",
+			"the machine that will run it, so the credential is delivered instead of printed")
+		positionals, err := parseAround(set, args[1:])
+		if err != nil {
+			return err
+		}
+		if len(positionals) != 1 {
+			return errors.New("module issue <module> --node <machine>")
+		}
+		module := positionals[0]
+		if *forNode == "" {
+			return errors.New("module issue needs --node: a module's account is sealed to the " +
+				"machine that runs it, and the mesh cannot read it back to print")
+		}
+
+		shelf, err := inv.Catalogue(ctx)
+		if err != nil {
+			return err
+		}
+		m, ok := shelf[module]
+		if !ok {
+			return fmt.Errorf("this mesh knows no module %q; `module add` it first", module)
+		}
+		if err := mayIssue(m); err != nil {
+			return err
+		}
+
+		// Which bus this mesh is on. A module gets a credential for exactly one, and the two are
+		// made in entirely different ways: on the bus the mesh runs on today an account is a
+		// management call, and on the bus being built it is a row the next composition writes into
+		// the server's user list (novox/hq design 25 §4).
+		busAddress, err := broker.BusAddress()
+		if err != nil {
+			return err
+		}
+		return issueOnTheNewBus(ctx, inv, m, *forNode, busAddress)
+
+	default:
+		return fmt.Errorf("module has no %q; it has add, check, list, moved, forget and issue", args[0])
+	}
+}
+
+func assignCommand(ctx context.Context, verb string, args []string) error {
+	// Several modules in one act (novox/hq ADR 0207): holders that depend on each other — the
+	// service manager and the package manager — can only go on, or come off, together.
+	if len(args) < 2 {
+		return fmt.Errorf("%s <node> <module> [<module>…]", verb)
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+
+	// The act itself is in acts.go, so the command API refuses exactly what this refuses
+	// (novox/hq ADR 0035). What differs between the surfaces is how the answer is printed.
+	act := assign
+	if verb == "unassign" {
+		act = unassign
+	}
+	said, err := act(ctx, open, args[0], args[1:]...)
+	if said != "" {
+		fmt.Println(said)
+	}
+	if err != nil {
+		fmt.Println()
+		return err
+	}
+	return nil
+}
+
+func settingsCommand(ctx context.Context, args []string) error {
+	if len(args) == 0 {
+		return errors.New("settings set <module> <file> [--node <node>], or settings clear <module> [--node <node>]")
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	set := flag.NewFlagSet("settings", flag.ContinueOnError)
+	node := set.String("node", "", "one machine, rather than the whole mesh")
+	positionals, err := parseAround(set, args[1:])
+	if err != nil {
+		return err
+	}
+
+	where := "the whole mesh"
+	if *node != "" {
+		where = *node
+	}
+
+	switch args[0] {
+	case "set":
+		if len(positionals) != 2 {
+			return errors.New("settings set <module> <settings.json | {…}> [--node <node>]")
+		}
+		// A file, or the values themselves when they begin with `{` — which is how the mesh's own
+		// `settings` tool passes them, having no file to hand over (novox/hq issue 198).
+		var raw []byte
+		if strings.HasPrefix(strings.TrimSpace(positionals[1]), "{") {
+			raw = []byte(positionals[1])
+		} else if raw, err = os.ReadFile(positionals[1]); err != nil {
+			return err
+		}
+		var values map[string]any
+		if err := json.Unmarshal(raw, &values); err != nil {
+			return fmt.Errorf("%s is not a settings file: %w", positionals[1], err)
+		}
+		if err := inv.SetSettings(ctx, *node, positionals[0], values); err != nil {
+			return err
+		}
+
+		var keys []string
+		for k := range values {
+			keys = append(keys, k)
+		}
+		sort.Strings(keys)
+		fmt.Printf("%s on %s: %s\n", positionals[0], where, strings.Join(keys, ", "))
+		fmt.Println("  run `push` to send it")
+		return nil
+
+	case "clear":
+		if len(positionals) != 1 {
+			return errors.New("settings clear <module> [--node <node>]")
+		}
+		if err := inv.ClearSettings(ctx, *node, positionals[0]); err != nil {
+			return err
+		}
+		fmt.Printf("%s on %s is back to what the module says\n", positionals[0], where)
+		return nil
+
+	default:
+		return fmt.Errorf("settings has no %q; it has set and clear", args[0])
+	}
+}
+
+// describeOffers says what a module provides, and marks the ones answered from anywhere in the
+// mesh — because "provides a database" and "provides a shell" are read the same way and mean
+// entirely different things about where the answer has to be.
+func describeOffers(offers []catalogue.Offer) string {
+	var out []string
+	for _, o := range offers {
+		if o.At() == catalogue.ScopeMesh {
+			out = append(out, o.Name+" (from anywhere in the mesh)")
+			continue
+		}
+		out = append(out, o.Name)
+	}
+	return strings.Join(out, ", ")
+}
+
+// pinCommand says which node a machine gets a provision from.
+//
+// Needed only when more than one could answer, and recordable before that -- a mesh with one
+// database should not change where an existing machine gets its data the day a second one
+// arrives.
+func pinCommand(ctx context.Context, args []string, setting bool) error {
+	if setting && len(args) != 4 {
+		return errors.New("pin <node> <provision> <from-node> <module>")
+	}
+	if !setting && len(args) != 2 {
+		return errors.New("unpin <node> <provision>")
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	if !setting {
+		if err := inv.UnpinProvision(ctx, args[0], args[1]); err != nil {
+			return err
+		}
+		fmt.Printf("%s is no longer told where to get %s from\n", args[0], args[1])
+		return nil
+	}
+	// The provider's node may be this same machine: two modules beside the consumer can both
+	// answer a provision, and then the module is the whole question (novox/hq #258).
+	if err := inv.PinProvision(ctx, args[0], args[1], args[2], args[3]); err != nil {
+		return err
+	}
+	fmt.Printf("%s gets %s from %s/%s\n", args[0], args[1], args[2], args[3])
+	fmt.Printf("  run `push %s` to send it\n", args[0])
+	return nil
+}
+
+// theBrokerSeat is the mesh-scoped seat the broker module claims (novox/hq ADR 0079: a
+// foundation seat is named after the server it guards).
+const theBrokerSeat = "mesh-broker"
+
+// brokerReachableAt is the broker's address as the given node can reach it.
+//
+// The genesis address (MESH_BROKER_ADDRESS) is the broker's public endpoint — right for a machine
+// that can dial it, wrong across the mesh, where the path is the overlay. A node that is ON the
+// overlay is given the overlay name of the node HOLDING the broker — the one assigned a module
+// claiming the mesh-broker seat — because that is where the broker is, whatever else the topology
+// says. Only when nothing holds the seat yet (genesis raised the broker as plumbing and no module
+// has adopted it) does the hub stand in, which is where the foundation is by convention.
+//
+// "On the overlay" is what `whereEveryoneIs` answers — a machine that RESOLVED the private network
+// — not "has an address", which is true of every placed machine and says nothing about
+// whether anything can reach it (novox/hq issue 059). A node not on the overlay — at genesis,
+// before any `overlay place`, which is when the builder's account is issued — keeps the genesis
+// address, so nothing about bring-up changes. This is issue 055, corrected by 059.
+func brokerReachableAt(ctx context.Context, inv *inventory.Inventory, known broker.Broker, node string) (string, error) {
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return "", err
+	}
+	onNetwork, err := whereEveryoneIs(ctx, inv, shelf)
+	if err != nil {
+		// No catalogue yet is genesis, and at genesis the genesis address is the right one.
+		return known.Address, nil
+	}
+	if onNetwork[node] == "" {
+		return known.Address, nil
+	}
+
+	// Where the broker is: the node whose assigned set includes a module claiming the seat.
+	holders := map[string]bool{}
+	for name, m := range shelf {
+		for _, c := range m.Claims {
+			if c.Name == theBrokerSeat && c.At() == catalogue.ScopeMesh {
+				holders[name] = true
+			}
+		}
+	}
+	brokerAt := ""
+	if len(holders) > 0 {
+		overlays, err := inv.Overlays(ctx)
+		if err != nil {
+			return "", err
+		}
+		for _, o := range overlays {
+			assigned, err := inv.Assigned(ctx, o.Name)
+			if err != nil {
+				continue
+			}
+			for _, a := range assigned {
+				if holders[a] && onNetwork[o.Name] != "" {
+					brokerAt = onNetwork[o.Name]
+				}
+			}
+		}
+	}
+	if brokerAt == "" {
+		// Nothing holds the seat (or its node is not on the overlay): the hub, where the
+		// foundation is by convention — and only a hub that is itself on the overlay, or the
+		// name we hand out routes nowhere.
+		overlays, err := inv.Overlays(ctx)
+		if err != nil {
+			return "", err
+		}
+		for _, o := range overlays {
+			if o.Hub && onNetwork[o.Name] != "" {
+				brokerAt = onNetwork[o.Name]
+			}
+		}
+	}
+	if brokerAt == "" {
+		return known.Address, nil
+	}
+	// A portless genesis address is a working configuration (amqps defaults to 5671), and
+	// silently keeping the public address would disable this whole path — so the port defaults
+	// rather than the fix dissolving (novox/hq issue 059).
+	_, port, err := net.SplitHostPort(known.Address)
+	if err != nil {
+		port = "5671"
+	}
+	return net.JoinHostPort(brokerAt, port), nil
+}
+
+// mayIssue says whether a module can be given a broker account. The account is delivered as the
+// module's own secret named broker; a module that declares none has nothing to read it with, and
+// an account nothing reads is an orphan on the bus (novox/hq 04-ISSUES/078). Said before the
+// account is made, not after.
+func mayIssue(m catalogue.Manifest) error {
+	if _, reads := m.OwnSecrets["broker"]; !reads {
+		return fmt.Errorf("%s declares no own secret named broker, so there is nowhere to deliver "+
+			"an account; a module that speaks on the bus declares \"own-secrets\": {\"broker\": <path>}", m.Module)
+	}
+	return nil
+}
+
+// issueOnTheNewBus gives an assigned module its credential on the bus being built.
+//
+// **Three things differ from a management call, and each is the point of the move.** The credential
+// is minted into the mesh's records and becomes usable at the next composition, so there is no
+// server to be reachable for this to work. The password travels beside the address rather than inside
+// it, because the runtime's contract already separates them and a credential embedded in a URL is one
+// that leaks into every log line that prints a connection. And the module's durable consumer is
+// derived from what it declared rather than declared by name, so a module cannot ask for delivery of
+// something it did not say it consumes.
+func issueOnTheNewBus(ctx context.Context, inv *inventory.Inventory, m catalogue.Manifest,
+	node, busAddress string) error {
+
+	user := broker.Principal{Kind: broker.KindModule, Node: node, Module: m.Module}.Username()
+	password, err := inv.MintBusPassword(ctx, inventory.BusUser{
+		Username: user, Kind: busKindOf(m.Module), Node: node, Module: m.Module,
+	})
+	if err != nil {
+		return err
+	}
+
+	// Where the module is told to find the bus, and what certificate it must present. The same pair
+	// a node is told, for the same reason: a mesh's bus presents its own certificate, in no public
+	// trust store, so an address alone fails at TLS.
+	known, err := broker.FromEnvironment()
+	if err != nil {
+		return fmt.Errorf("cannot deliver a credential without knowing where the bus is: %w", err)
+	}
+	reachable, err := brokerReachableAt(ctx, inv, known, node)
+	if err != nil {
+		return err
+	}
+	return issueWith(ctx, inv, m, node, busAddress, known, reachable, user, password)
+}
+
+// busKindOf is what a module's bus user is recorded as: the node's tool runtime where the module is
+// the runtime (novox/hq ADR 0175), a module otherwise. The username is the same either way — the
+// runtime is issued through this same path — and the kind is what a reader of the records sees.
+func busKindOf(module string) string {
+	if module == catalogue.RuntimeModule {
+		return inventory.BusNodeTools
+	}
+	return inventory.BusModule
+}
+
+// issueWith is the delivery half: the minted password sealed to the machine as the module's broker
+// secret, and the module's consumer created where the bus can be reached. Split from the minting
+// so the move can issue every module against a bus whose address it worked out itself
+// (`rollout mint`, design 28 task 5.2) rather than the one in this process's environment.
+func issueWith(ctx context.Context, inv *inventory.Inventory, m catalogue.Manifest,
+	node, busAddress string, known broker.Broker, reachable, user, password string) error {
+	// The seats this module claims, with the verbs each promises (novox/hq ADR 0159): the runtime
+	// serves a claimed seat's verbs with its tools of the same name, and the bus admits only the
+	// holder's subscription — so the runtime tries each claim and the grant decides. Written here
+	// because this file is the one thing the mesh writes that the runtime reads before it speaks.
+	claims, err := claimsFor(ctx, inv, m)
+	if err != nil {
+		return err
+	}
+	held, err := json.Marshal(struct {
+		URL         string        `json:"url"`
+		Fingerprint string        `json:"fingerprint,omitempty"`
+		Node        string        `json:"node"`
+		Module      string        `json:"module"`
+		User        string        `json:"user"`
+		Password    string        `json:"password"`
+		Claims      []seatClaimed `json:"claims,omitempty"`
+	}{
+		URL: "nats://" + reachable, Fingerprint: known.Fingerprint,
+		Node: node, Module: m.Module, User: user, Password: password, Claims: claims,
+	})
+	if err != nil {
+		return err
+	}
+	if err := inv.AcceptSecretForModule(ctx, node, m.Module, "broker", string(held)); err != nil {
+		return err
+	}
+
+	// And how it hears what it consumes. Derived from its declaration, and only when it declared
+	// something: a module that consumes nothing needs no consumer, and creating one would be a
+	// durable subscription nobody reads.
+	if consumer, needed := broker.ConsumerFor(broker.Principal{
+		Kind: broker.KindModule, Node: node, Module: m.Module,
+		Emits: m.EmitsAll(), Consumes: m.Consumes, Serves: m.Tools,
+	}); needed {
+		if busAddress == "" {
+			fmt.Printf("  %s consumes; its consumer is created when the bus is reachable (`push`, then "+
+				"`rollout mint` again is harmless)\n", m.Module)
+		} else {
+			js, err := broker.Dial(busAddress)
+			if err != nil {
+				return fmt.Errorf("the credential is minted and the mesh cannot reach the bus to create "+
+					"how %s hears what it consumes: %w", m.Module, err)
+			}
+			defer js.Close()
+			if err := js.EnsureConsumer(consumer); err != nil {
+				return err
+			}
+		}
+	}
+
+	fmt.Printf("bus user %s minted for %s, scoped to what it emits and consumes\n", user, m.Module)
+	fmt.Printf("  sealed to %s. It arrives with the next push — `push %s` to send it\n", node, node)
+	fmt.Printf("  and it works once the bus has been told: the user list is composed into the " +
+		"machine holding mesh-broker\n")
+	return nil
+}
+
+// whereItComesFrom is the provenance a module handed over by hand records, and what a record must
+// say to be worth anything later.
+//
+// **A module is a repository and a directory inside it** (novox/hq ADR 0069). A record carrying only
+// the repository names a module.json at its root, so every later build of it looks in the wrong
+// place — nine modules on this mesh were registered that way and none of them could be rebuilt
+// (2026-09-28). The directory cannot be checked from here, because the control plane does not clone;
+// what can be checked is that the record is whole.
+func whereItComesFrom(repository, ref, commit, path string, self bool) (inventory.Source, error) {
+	// Provenance together or not at all. A source with no commit cannot be compared against
+	// anything, so it would record where the module came from and still never be able to say the
+	// mesh is behind it — which is the one thing recording it is for.
+	if (repository == "") != (commit == "") {
+		return inventory.Source{}, errors.New("--source and --commit go together: a source with " +
+			"no commit cannot be compared against anything, and a commit with no source has " +
+			"nothing to be compared with")
+	}
+	// A directory or a forge with no repository is half a location, and the half it keeps is the
+	// half nothing can be found with.
+	if repository == "" && (path != "" || self) {
+		return inventory.Source{}, errors.New("--path and --self say where inside a source and " +
+			"which forge holds it, so they need --source: without one there is nothing for them " +
+			"to be part of")
+	}
+	from := inventory.Source{Repository: repository, Ref: ref, BuiltFrom: commit, Path: path}
+	if self {
+		if err := onASeat(repository); err != nil {
+			return inventory.Source{}, err
+		}
+		from.Seat = gitSeat
+	}
+	return from, nil
+}
+
+// namesNoInstallation is the mesh refusing a definition that names an installation, at the moment
+// it would enter the catalogue (novox/hq ADR 0112, ADR 0155). `module check` says the same thing
+// earlier, where the author is; this is the last moment the mesh can still say no, and a
+// definition that got past the check — written elsewhere, or checked by nobody — is refused here
+// in the same words. A name meant on purpose is declared with its reason and passes.
+func namesNoInstallation(m catalogue.Manifest) error {
+	named := catalogue.InstallationProblems(m)
+	if len(named) == 0 {
+		return nil
+	}
+	return fmt.Errorf("%s names an installation, and a definition names none — declare a name meant "+
+		"on purpose under %s with its reason, or take it out:\n  - %s",
+		m.Module, catalogue.NamesOnPurpose, strings.Join(named, "\n  - "))
+}
+
+// seatClaimed is one seat a module claims, as its runtime needs it: the name, the scope (a
+// node-scoped seat's verb carries the machine, design 33 §4) and the verbs the seat promises.
+type seatClaimed struct {
+	Seat   string   `json:"seat"`
+	Scope  string   `json:"scope"`
+	Serves []string `json:"serves,omitempty"`
+}
+
+// claimsFor joins a module's claims with the seats' protocols from the mesh's records.
+func claimsFor(ctx context.Context, inv *inventory.Inventory, m catalogue.Manifest) ([]seatClaimed, error) {
+	if len(m.Claims) == 0 {
+		return nil, nil
+	}
+	seats, err := inv.Seats(ctx)
+	if err != nil {
+		return nil, err
+	}
+	byName := map[string]catalogue.Seat{}
+	for _, s := range seats {
+		byName[s.Name] = s
+	}
+	var out []seatClaimed
+	for _, c := range m.Claims {
+		claimed := seatClaimed{Seat: c.Name, Scope: c.At()}
+		if s, known := byName[c.Name]; known {
+			claimed.Scope = s.Scope
+			// The verbs the runtime serves for the seat: the claim's own when it names them
+			// (ADR 0160), else every verb the seat promises, which its tools then answer.
+			claimed.Serves = c.ServesFor(catalogue.Manifest{Tools: catalogue.VerbNames(s.Serves)})
+		}
+		out = append(out, claimed)
+	}
+	return out, nil
+}
+
+// declaresTools is whether a module answers tools wherever it runs (novox/hq ADR 0197): it names
+// tools of its own, its build delivers a bundle the node's runtime serves, or it claims a seat
+// whose verbs it serves. A module with none is never expected to announce anything.
+func declaresTools(m catalogue.Manifest) bool {
+	if len(m.Tools) > 0 {
+		return true
+	}
+	for _, b := range m.Bundles {
+		if len(b.Loads) > 0 {
+			return true
+		}
+	}
+	for _, c := range m.Claims {
+		if len(c.Serves) > 0 {
+			return true
+		}
+	}
+	return false
+}

cmd/mesh-controller/modules_test.go

@@ -0,0 +1,26 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// `module issue` makes a broker account and delivers it as the module's own secret named broker.
+// A module that declares none is refused before the account exists, so the bus never carries an
+// account nothing reads (novox/hq 04-ISSUES/078).
+func TestAModuleWithNoBrokerSecretCannotBeIssued(t *testing.T) {
+	err := mayIssue(catalogue.Manifest{Module: "step-ca", OwnSecrets: catalogue.OwnSecrets{"password": {Path: "/run/password"}}})
+	if err == nil {
+		t.Fatal("a module with no broker own secret was issued an account")
+	}
+	for _, want := range []string{"step-ca", "broker", "own-secrets"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Errorf("the refusal does not say %q: %v", want, err)
+		}
+	}
+	if err := mayIssue(catalogue.Manifest{Module: "redis", OwnSecrets: catalogue.OwnSecrets{"broker": {Path: "/run/broker"}}}); err != nil {
+		t.Errorf("a module declaring its broker secret was refused: %v", err)
+	}
+}

cmd/mesh-controller/network.go

@@ -0,0 +1,732 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"os"
+	"sort"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/overlay"
+)
+
+// the private network: who is on it, where, and what they are called.
+//
+// Split out of main.go, which had reached 2,769 lines because appending was always the
+// cheapest next step. That is how novox/hq ADR 0001 records `hal/sdk` reaching 34,636:
+// nothing in it was wrong, and no one edit was the one that should have been a new file.
+
+// DefaultOverlayCIDR is the range the mesh allocates from when nothing says another.
+const DefaultOverlayCIDR = "10.42.0.0/16"
+
+// overlayRange is the range the mesh allocates node addresses from.
+//
+// **The adopted tunnel's range first** (novox/hq ADR 0105): a hub that took over the tunnel it
+// found is at that tunnel's address, its peers are at theirs, and every node's address is
+// composed from the same range — the hub's, and every binding, hosts entry and endpoint derived
+// from it. Those are readers of this; none of them stores the range. Without an adopted tunnel,
+// the range genesis was told, or the default.
+func overlayRange(ctx context.Context, inv *inventory.Inventory) (string, error) {
+	tunnel, _, adopted, err := inv.AdoptedTunnel(ctx)
+	if err != nil {
+		return "", err
+	}
+	if adopted {
+		return tunnel.Range, nil
+	}
+	if v := strings.TrimSpace(os.Getenv(OverlayCIDRVar)); v != "" {
+		return v, nil
+	}
+	return DefaultOverlayCIDR, nil
+}
+
+func overlayCommand(ctx context.Context, args []string) error {
+	if len(args) == 0 {
+		return errors.New("overlay place <node> [flags], overlay name <address> <name>, or overlay show")
+	}
+	// Answered before anything is opened. A message about which command to use should not need a
+	// database to say so, and needing one turns a redirect into a connection error.
+	if args[0] == "push" {
+		return errors.New("`overlay push` is now `push`, which sends a node its network AND " +
+			"what its assignments resolve to — the two are computed from one picture of the " +
+			"mesh, and sending them separately would let them disagree")
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	switch args[0] {
+	case "place":
+		return overlayPlace(ctx, inv, args[1:])
+	case "show":
+		return overlayShow(ctx, open)
+	case "name":
+		return overlayName(ctx, inv, args[1:])
+
+	default:
+		return fmt.Errorf("overlay has no %q; it has place, name and show", args[0])
+	}
+}
+
+// overlayName is the operator saying which machine a carried address is (novox/hq issue 112),
+// so the mesh answers for its name until the machine enrols and verifies it.
+func overlayName(ctx context.Context, inv *inventory.Inventory, args []string) error {
+	if len(args) != 2 {
+		return errors.New("overlay name <carried-address> <node-name>")
+	}
+	address, name := args[0], args[1]
+	if err := inv.NamePeer(ctx, address, name); err != nil {
+		return err
+	}
+	fmt.Printf("the peer at %s is %s until it enrols — the mesh answers for %s.<suffix> from the "+
+		"operator's word, and enrolment under this key must use this name\n", address, name, name)
+	return nil
+}
+
+func overlayPlace(ctx context.Context, inv *inventory.Inventory, args []string) error {
+	if len(args) == 0 {
+		return errors.New(
+			"overlay place <node> [--endpoint host:port] [--site name] [--hub], or --nothing")
+	}
+	node := args[0]
+
+	set := flag.NewFlagSet("overlay place", flag.ContinueOnError)
+	endpoint := set.String("endpoint", "", "where this node can be dialled, or empty for nowhere")
+	site := set.String("site", "", "where this machine physically is, or empty if it roams")
+	hub := set.Bool("hub", false, "this node is the hub every other routes through")
+	nothing := set.Bool("nothing", false,
+		"place it with nothing set: not dialable, no site, not the hub")
+	if err := set.Parse(args[1:]); err != nil {
+		return err
+	}
+
+	// **All three are declared together, so saying nothing took all three away.** The sibling of
+	// `node public-domain`: `overlay place anchor` reads like it places the node it names, and it
+	// silently unset the endpoint every other machine dials, the site it is in, and the hub if it
+	// was the hub — every path through it going with them, at the moment somebody was trying to
+	// look at it.
+	//
+	// A placement with nothing set is a real thing to want — a machine that roams and opens every
+	// path itself is exactly that — so it keeps a way to say so, by name.
+	if set.NFlag() == 0 {
+		return fmt.Errorf("overlay place %s was given nothing to place it with, and all three are "+
+			"declared together — it would take away the endpoint other machines dial %s at, its "+
+			"site, and the hub if it is the hub. Say --endpoint/--site/--hub, or --nothing if that "+
+			"is what you meant", node, node)
+	}
+	if *nothing && (*endpoint != "" || *site != "" || *hub) {
+		return fmt.Errorf("give %s a placement or --nothing, not both: they say opposite things "+
+			"and the mesh will not choose between them", node)
+	}
+
+	// One hub per mesh, refused rather than last-write-wins: with two flagged, which one the
+	// graph and the broker address pick is order-dependent — the silent-election fault ADR 0007
+	// exists to avoid, one flag over (novox/hq issue 059). Moving the hub is explicit: re-place
+	// the old one without --hub first.
+	if *hub {
+		placed, err := inv.Overlays(ctx)
+		if err != nil {
+			return err
+		}
+		for _, p := range placed {
+			if p.Hub && p.Name != node {
+				return fmt.Errorf("%s is already the hub; a mesh has one. Re-place %s without "+
+					"--hub first if the hub is moving", p.Name, p.Name)
+			}
+		}
+	}
+
+	// A hub that took over a tunnel listens on that tunnel's port — it is what the peers dial, and
+	// the reason the port is worth having (novox/hq ADR 0105). An endpoint on another port would
+	// have the mesh's interface up where no peer is listening for it.
+	found, err := inv.NodeByName(ctx, node)
+	if err != nil {
+		return err
+	}
+	var tunnel inventory.Tunnel
+	adoptsTunnel := false
+	if *hub && found.Adopted {
+		if t, err := inv.TunnelOf(ctx, node); err == nil {
+			tunnel = t
+			placed, _ := inv.Overlays(ctx)
+			for _, o := range placed {
+				if o.Name == node && o.Key == t.PublicKey {
+					adoptsTunnel = true
+				}
+			}
+		} else if !errors.Is(err, inventory.ErrNoTunnel) {
+			return err
+		}
+	}
+	if adoptsTunnel {
+		if port := portOfEndpoint(*endpoint); port != strconv.Itoa(tunnel.Port) {
+			return fmt.Errorf("%s takes over the tunnel it found on %s, which listens on port %d, and "+
+				"its endpoint %q names another port: the peers dial the tunnel's port, so the hub's "+
+				"endpoint must be on it", node, tunnel.Interface, tunnel.Port, *endpoint)
+		}
+	}
+
+	// Declared, all three. The address is evidence of reachability and is not the fact, and hub
+	// election by address prefix fails silently (novox/hq ADR 0007).
+	if err := inv.SetPlace(ctx, node, *endpoint, *site, *hub, ""); err != nil {
+		return err
+	}
+	cidr, err := overlayRange(ctx, inv)
+	if err != nil {
+		return err
+	}
+	address, err := inv.AssignAddress(ctx, found.ID, cidr)
+	if err != nil {
+		return err
+	}
+
+	fmt.Printf("%s is at %s on the overlay\n", node, address)
+	fmt.Println("  credentials issued for it before this placement keep their old broker address —" +
+		" `module issue` them again and push (novox/hq issue 059)")
+	switch {
+	case adoptsTunnel:
+		fmt.Printf("  the hub — it takes over the tunnel it found on %s: range %s, port %d, "+
+			"%d peer(s) carried until they enrol\n", tunnel.Interface, tunnel.Range, tunnel.Port,
+			len(tunnel.Peers))
+	case *hub:
+		fmt.Println("  the hub — every node not sharing a site routes through it")
+	case *endpoint == "":
+		fmt.Println("  not dialable — it opens every path itself")
+	}
+	if *site != "" {
+		fmt.Printf("  at %s, so it peers directly with anything else there\n", *site)
+	}
+	return nil
+}
+
+// network builds the private network over the machines that resolved the module for it.
+//
+// Not over every node the mesh knows. **A machine is on the private network because it was given
+// the module**, and one that was not is absent from every peer list and from the names — which is
+// the only thing "not on the network" can mean. Until this, having an address was enough, and
+// there was no way to keep a machine off.
+//
+// Every node at once, which is the whole reason this is the control plane's work: a peer list is
+// derived from all the others, so no node could compute its own.
+func network(ctx context.Context, inv *inventory.Inventory, on map[string]bool,
+	refused map[string]string) (*overlay.Generator, error) {
+	places, err := inv.Overlays(ctx)
+	if err != nil {
+		return nil, err
+	}
+	// The tunnels adopted nodes take over, and the peers the hub's carries (novox/hq ADR 0105).
+	tunnels, err := inv.Tunnels(ctx)
+	if err != nil {
+		return nil, err
+	}
+	carried, err := inv.CarriedPeers(ctx)
+	if err != nil {
+		return nil, err
+	}
+	nodes := make([]overlay.Node, 0, len(places))
+	for _, p := range places {
+		if !on[p.Name] {
+			continue
+		}
+		n := overlay.Node{
+			Name: p.Name, Key: p.Key, Endpoint: p.Endpoint,
+			Site: p.Site, Hub: p.Hub, Address: p.Address,
+		}
+		if t, takes := tunnels[p.Name]; takes && t.NodeAdopted {
+			// Only an adopted node is told to take the found unit over: on a converged one there
+			// is nothing found to keep, and the host refuses the field. The range and the carried
+			// peers do not depend on the mode; the takeover does.
+			//
+			// **Refused, not composed, when the hub's record disagrees with the tunnel.** A
+			// declaration that stopped the found unit and raised the mesh's interface on another
+			// port or address would leave every peer dark while reporting the tunnel taken — so a
+			// hub placed before it took the tunnel over (or at the wrong port) is named here, and
+			// nothing is sent until it is re-placed.
+			if wrong := disagrees(p, t.Tunnel); wrong != "" {
+				return nil, fmt.Errorf("%s takes over the tunnel on %s and its placement disagrees with it: %s. "+
+					"Re-place it — `overlay place %s --hub --endpoint <host>:%d …` — and push again; "+
+					"nothing was composed", p.Name, t.Interface, wrong, p.Name, t.Port)
+			}
+			n.TakesOver = &overlay.TakeOver{Interface: t.Interface, Unit: t.Unit, Config: t.Config, Port: t.Port, MTU: t.MTU}
+		}
+		if p.Hub {
+			for _, c := range carried {
+				n.Carried = append(n.Carried, overlay.Carried{Key: c.PublicKey, Address: c.Address})
+			}
+		}
+		nodes = append(nodes, n)
+	}
+	if len(nodes) == 0 {
+		// Nobody was given it. An empty network is a legitimate mesh, not a broken one, so this
+		// answers rather than refusing -- Compute would refuse for want of a hub, and reporting
+		// "no hub" to somebody who never asked for a network would be a lie about the cause.
+		return overlay.Empty(), nil
+	}
+	cidr, err := overlayRange(ctx, inv)
+	if err != nil {
+		return nil, err
+	}
+	// No registry trust is composed here any more: the container runtime's module states it, told
+	// where the store is reached by ${seat:mesh-artifact-store:reach} (novox/hq ADR 0222, issue 190).
+	g, err := overlay.From(nodes, cidr, "")
+	if err != nil && len(refused) > 0 {
+		// The network is missing something, and some machines could not be resolved at all. Those
+		// are almost always the same fact: a node that does not resolve contributes nothing, so
+		// reporting "no hub" would name a consequence and hide the cause.
+		var who []string
+		for name, why := range refused {
+			who = append(who, fmt.Sprintf("  %s: %s", name, why))
+		}
+		sort.Strings(who)
+		return nil, fmt.Errorf("%w\n\nand %d node(s) could not be resolved at all, which is "+
+			"probably why:\n%s", err, len(refused), strings.Join(who, "\n"))
+	}
+	return g, err
+}
+
+// graph is the whole mesh's network, for showing it.
+func graph(ctx context.Context, open *stores) ([]overlay.Node, overlay.Graph, error) {
+	inv := open.inventory
+	on, refused, err := whoResolves(ctx, open, overlay.Requirement)
+	if err != nil {
+		return nil, nil, err
+	}
+	g, err := network(ctx, inv, on, refused)
+	if err != nil {
+		return nil, nil, err
+	}
+	return g.Nodes(), g.Graph(), nil
+}
+
+// whoResolves is the machines whose resolution answers a requirement, and why the others did not.
+//
+// By what a module **provides**, not by its name. WireGuard is one way to have a private network
+// and there could be others, so a machine is on the network because something it runs provides
+// one — asking for a particular module by name would be the mistake this whole mechanism exists
+// to avoid.
+//
+// Resolved rather than read from the assignment table, because a module can arrive by being
+// required by something else, and a machine that needs the private network to do its job is on it
+// for the same reason as one that was handed it directly.
+func whoResolves(ctx context.Context, open *stores, requirement string) (
+	map[string]bool, map[string]string, error) {
+	inv := open.inventory
+	nodes, err := inv.Nodes(ctx)
+	if err != nil {
+		return nil, nil, err
+	}
+	on := map[string]bool{}
+	// Why a node could not be resolved, kept rather than raised: one broken node must not stop
+	// the rest being described, and whoever is rendering that node will raise it themselves.
+	refused := map[string]string{}
+	for _, n := range nodes {
+		plan, _, err := planFor(ctx, open, n.Name)
+		switch {
+		case unresolvable(err):
+			refused[n.Name] = err.Error()
+			continue
+		case err != nil:
+			// Not a node that does not resolve — a question that went unanswered. Recording it as a
+			// refusal would take the machine off the private network, and the generator that reads
+			// this would then write a roster and a filter without it (novox/hq 04-ISSUES/152).
+			return nil, nil, fmt.Errorf("whether %s answers %q cannot be read: %w",
+				n.Name, requirement, err)
+		}
+		for _, m := range plan.Modules {
+			for _, offered := range m.Offers() {
+				if offered == requirement {
+					on[n.Name] = true
+				}
+			}
+		}
+	}
+	return on, refused, nil
+}
+
+// rendering is everything a declaration needs, computed over the whole mesh.
+func generators(ctx context.Context, open *stores) (
+	map[string]catalogue.Generator, error) {
+	inv := open.inventory
+	on, refused, err := whoResolves(ctx, open, overlay.Addressing)
+	if err != nil {
+		return nil, err
+	}
+	net, err := network(ctx, inv, on, refused)
+	if err != nil {
+		return nil, err
+	}
+	// **One generator now.** Two more used to sit beside it — the names and a resolver's zone
+	// file — as modules that ran nothing. Both are facts a module asks for in its manifest
+	// (`facts:` — catalogue.FactsInto), computed from the same machines this sees: the ones on the
+	// private network, because a name for a machine not on it would resolve to an address nothing
+	// can reach.
+	return map[string]catalogue.Generator{
+		overlay.Name: net,
+	}, nil
+}
+
+func overlayShow(ctx context.Context, open *stores) error {
+	nodes, computed, err := graph(ctx, open)
+	if err != nil {
+		return err
+	}
+	if len(nodes) == 0 {
+		// Not "this mesh has no nodes", which it said until the network became a module and was
+		// then a lie about the cause: a mesh can have every node it will ever have and nobody on
+		// the private network, because nobody asked for one.
+		fmt.Printf("nobody is on the private network — assign %s to put a machine on it\n",
+			overlay.Name)
+		return nil
+	}
+
+	// The tunnel the hub took over, if any, and the peers carried from it (novox/hq ADR 0105):
+	// listed apart from the nodes, because they are peers of the tunnel and not nodes of the
+	// mesh until they enrol — and once one has, it is listed as the node it became.
+	tunnel, hubName, adopted, err := open.inventory.AdoptedTunnel(ctx)
+	if err != nil {
+		return err
+	}
+	carried, err := open.inventory.CarriedPeers(ctx)
+	if err != nil {
+		return err
+	}
+	for _, n := range nodes {
+		place := n.Address
+		if place == "" {
+			// Said, not skipped. A node with no place is a node with no network, and it should
+			// be visible here rather than quietly absent from a list of who is on it.
+			place = "no address — run `overlay place`"
+		}
+		fmt.Printf("%-16s %-14s", n.Name, place)
+		switch {
+		case n.Hub && adopted:
+			fmt.Printf("  hub — over the tunnel it took over on %s (range %s, port %d)",
+				tunnel.Interface, tunnel.Range, tunnel.Port)
+		case n.Hub && hubName == n.Name && tunnel.Interface != "":
+			fmt.Printf("  hub — found a tunnel on %s and did NOT take it over: its key is not the tunnel's; "+
+				"`mesh-host overlay take --tunnel %s` on the machine takes it", tunnel.Interface, tunnel.Interface)
+		case n.Hub:
+			fmt.Print("  hub — found no tunnel; if the machine runs the predecessor's, " +
+				"`mesh-host overlay take --tunnel <iface>` there adopts it (novox/hq ADR 0105)")
+		case !n.Reachable():
+			fmt.Print("  not dialable")
+		}
+		if n.Site != "" {
+			fmt.Printf("  at %s", n.Site)
+		}
+		if n.TakesOver != nil && !n.Hub {
+			fmt.Printf("  takes over %s", n.TakesOver.Interface)
+		}
+		fmt.Println()
+		for _, p := range computed[n.Name] {
+			fmt.Printf("    → %-14s %-18s %s\n", p.Name, p.Allowed, p.Why)
+		}
+	}
+	if len(carried) > 0 {
+		fmt.Printf("\npeers of the tunnel %s took over — not nodes of the mesh until they enrol:\n", hubName)
+		for _, c := range carried {
+			state := "not yet enrolled"
+			if c.EnrolledAs != "" {
+				state = "enrolled as " + c.EnrolledAs + ", which keeps this address"
+			}
+			fmt.Printf("  %-16s %-14s %s\n", overlay.CarriedName(c.PublicKey), c.Address, state)
+		}
+	}
+	return nil
+}
+
+// disagrees says how a node's placement differs from the tunnel it takes over — its address not
+// the tunnel's, its endpoint not on the tunnel's port — or nothing when both agree.
+func disagrees(p inventory.Overlay, t inventory.Tunnel) string {
+	var wrong []string
+	want := t.Address
+	if i := strings.Index(want, "/"); i >= 0 {
+		want = want[:i]
+	}
+	if p.Address != want {
+		wrong = append(wrong, fmt.Sprintf("its address is %s and the tunnel's is %s", orNothing(p.Address), want))
+	}
+	if p.Reachable() && portOfEndpoint(p.Endpoint) != strconv.Itoa(t.Port) {
+		wrong = append(wrong, fmt.Sprintf("its endpoint %s is not on the tunnel's port %d", p.Endpoint, t.Port))
+	}
+	return strings.Join(wrong, "; ")
+}
+
+func orNothing(s string) string {
+	if s == "" {
+		return "unset"
+	}
+	return s
+}
+
+// portOfEndpoint is the port in host:port, or empty.
+func portOfEndpoint(endpoint string) string {
+	if i := strings.LastIndex(endpoint, ":"); i >= 0 {
+		return endpoint[i+1:]
+	}
+	return ""
+}
+
+// SilentFor is how long a node may be quiet before the mesh says so.
+//
+// A node speaks every minute, so three of them missed is a gap rather than a slow one. The number
+// is not the point — being able to say "out of touch" at all is, and nothing could before.
+const SilentFor = 3 * time.Minute
+
+// whereEveryoneIs is each machine's name on the private network, for the ones on it.
+//
+// **Resolved without consulting the rest of the mesh**, and that is not an optimisation. Every
+// other path here answers a question about one node by resolving the others; this one is called
+// *from* that path, so doing the same would not terminate — which it did not, for two minutes,
+// until it was run.
+//
+// An unchecked resolution is exactly right for the question anyway. Whether a machine is on the
+// private network depends on what it was assigned and what that requires, both of which are local
+// facts. What it takes *from* other machines does not change the answer.
+//
+// The distinction that matters is kept: a machine absent from the network module's own view is
+// absent here, so "has an address" is not mistaken for "is reachable" — which it was, before the
+// network became something a machine is given.
+func whereEveryoneIs(ctx context.Context, inv *inventory.Inventory,
+	shelf map[string]catalogue.Manifest) (map[string]string, error) {
+
+	if shelf == nil {
+		// Refused rather than answered. Being on the private network is a conclusion about what a
+		// node resolves to, so with no catalogue nothing resolves and the honest answer is
+		// "nobody" — which is wrong, indistinguishable from a mesh with no overlay, and refused
+		// every certificate the mesh was asked for while saying the machine was on no network.
+		return nil, errors.New(
+			"asked where everyone is without the catalogue, which cannot be answered")
+	}
+	places, err := onTheNetwork(ctx, inv, shelf)
+	if err != nil {
+		return nil, err
+	}
+	out := map[string]string{}
+	for _, p := range places {
+		out[p.Name] = overlay.InternalName(p.Name)
+	}
+	return out, nil
+}
+
+// onTheNetwork is every placed machine that resolves the private network — has an address AND
+// runs what puts it there. "Has an address" alone was true of every placed machine and told you
+// nothing about whether anything could reach it; a name written for such a machine resolves to
+// an address that does not answer, and a connection to it hangs (novox/hq issue 079).
+func onTheNetwork(ctx context.Context, inv *inventory.Inventory,
+	shelf map[string]catalogue.Manifest) ([]inventory.Overlay, error) {
+	places, err := inv.Overlays(ctx)
+	if err != nil {
+		return nil, err
+	}
+	// **With the seat holders on record**, or a machine running the next holder of a seat beside
+	// the current one resolves as two holders, is refused, and drops out of the map — taking the
+	// address every other machine composes for what it offers (novox/hq ADR 0131). Found live:
+	// the control node vanished from the private network the moment the new bus was assigned
+	// beside the old one.
+	holdings, err := inv.Holdings(ctx)
+	if err != nil {
+		return nil, err
+	}
+	var out []inventory.Overlay
+	for _, p := range places {
+		if p.Address == "" {
+			continue
+		}
+		assigned, err := inv.Assigned(ctx, p.Name)
+		if err != nil || len(assigned) == 0 {
+			continue
+		}
+		caps, _ := inv.ProfileOf(ctx, p.Name)
+		got, err := catalogue.Resolve(shelf, assigned,
+			catalogue.Node{Name: p.Name, Site: p.Site, Capabilities: caps},
+			catalogue.World{Unchecked: true, Holdings: holdings})
+		if err != nil {
+			// Said, not skipped in silence: a machine dropped here loses its address, and every
+			// plan that names it fails in another module's words (novox/hq issue 188).
+			fmt.Fprintf(os.Stderr, "%s is not counted as on the network: it does not resolve: %v\n", p.Name, err)
+			continue
+		}
+		for _, m := range got.Modules {
+			for _, offered := range m.Offers() {
+				if offered == overlay.Requirement {
+					out = append(out, p)
+				}
+			}
+		}
+	}
+	return out, nil
+}
+
+// onThePrivateNetwork is every node's address on the private network, sorted — the same set
+// the names and the resolver mean by it (onTheNetwork), so a rule saying "from the mesh" admits
+// exactly the machines the mesh names.
+//
+// A node with no address is left out rather than rendered as an empty source: an empty entry in a
+// source set is a syntax error in the rule file, and a rule file that does not load leaves the
+// node filtering whatever it was filtering before -- the one outcome worse than a wrong rule,
+// because nothing reports it.
+func onThePrivateNetwork(ctx context.Context, inv *inventory.Inventory,
+	shelf map[string]catalogue.Manifest) ([]string, error) {
+	places, err := onTheNetwork(ctx, inv, shelf)
+	if err != nil {
+		return nil, err
+	}
+	var out []string
+	for _, p := range places {
+		if strings.TrimSpace(p.Address) != "" {
+			out = append(out, p.Address)
+		}
+	}
+	sort.Strings(out)
+	return out, nil
+}
+
+// namesInTheMesh is every machine's internal name and the address behind it — every machine
+// that is on the private network, the same set the resolver means by that.
+//
+// A machine that is not has no name: writing one that resolves to nothing is worse than not
+// writing it, because a connection to an address that does not answer hangs where a name that
+// does not resolve fails at once and says so. A machine placed on the overlay but not running
+// the module that puts it there is exactly that (novox/hq issue 079).
+func namesInTheMesh(ctx context.Context, inv *inventory.Inventory,
+	shelf map[string]catalogue.Manifest) (map[string]string, error) {
+	places, err := onTheNetwork(ctx, inv, shelf)
+	if err != nil {
+		return nil, err
+	}
+	out := map[string]string{}
+	for _, p := range places {
+		out[overlay.InternalName(p.Name)] = p.Address
+	}
+	// And the carried peers the operator has named (novox/hq issue 112): machines the
+	// predecessor's resolver answers for and the mesh routes to, known by name on the operator's
+	// word until they enrol — at which point enrolment verifies the name and the node's own
+	// entry takes over above. A name the predecessor answers for must keep resolving until the
+	// machine behind it is a node; without these, taking the resolver silences three machines.
+	carried, err := inv.CarriedPeers(ctx)
+	if err != nil {
+		return nil, err
+	}
+	for _, p := range carried {
+		if p.Named == "" || p.EnrolledAs != "" {
+			continue
+		}
+		if _, taken := out[overlay.InternalName(p.Named)]; taken {
+			continue // a node of the mesh owns the name; the stale statement loses
+		}
+		out[overlay.InternalName(p.Named)] = p.Address
+	}
+	return out, nil
+}
+
+// artifactStoreOnNetwork is the machine on this network that offers the artifact store, and the
+// port THAT MACHINE put it on — the node's setting when it was given one (novox/hq ADR 0100,
+// 04-ISSUES/102), the mesh's assignment when it made one, and the manifest's own number only when
+// neither says anything. Read exactly as a consumer's binding is, because the trust a machine
+// writes for the store and the address it pulls from are the same fact as what a consumer is told.
+//
+// A lookup failure is an error, never "not found": collapsing the two composed a declaration
+// without the trust whenever the inventory hiccuped, delivered by a push that reported success —
+// and nothing recomposed the machine until the next push. "No store" must mean the mesh has none,
+// not that the question went unanswered.
+func artifactStoreOnNetwork(ctx context.Context, inv *inventory.Inventory,
+	on map[string]bool) (node, port string, found bool, err error) {
+
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return "", "", false, fmt.Errorf("reading the catalogue: %w", err)
+	}
+	providers := map[string]catalogue.Manifest{}
+	for name, m := range shelf {
+		if _, offers := m.Serves[catalogue.ArtifactStoreProvision]; offers {
+			providers[name] = m
+		}
+	}
+	if len(providers) == 0 {
+		return "", "", false, nil
+	}
+	// In a stated order, so two machines offering it would always answer the same one.
+	machines := make([]string, 0, len(on))
+	for machine := range on {
+		machines = append(machines, machine)
+	}
+	sort.Strings(machines)
+	for _, machine := range machines {
+		assigned, err := inv.Assigned(ctx, machine)
+		if err != nil {
+			return "", "", false, fmt.Errorf("reading what %s is assigned: %w", machine, err)
+		}
+		for _, a := range assigned {
+			m, offers := providers[a]
+			if !offers {
+				continue
+			}
+			serves, err := servedOnNode(ctx, inv, machine, m, catalogue.ArtifactStoreProvision)
+			if err != nil {
+				return "", "", false, fmt.Errorf("reading where %s puts the artifact store: %w", machine, err)
+			}
+			if p, ok := serves["port"]; ok {
+				return machine, fmt.Sprintf("%v", p), true, nil
+			}
+		}
+	}
+	return "", "", false, nil
+}
+
+// artifactStoreAddress is the artifact store as `forNode` reaches it: `<node>.internal:<port>`
+// over the private network, or — when nothing is on the network yet — `127.0.0.1:<port>` for the
+// node that holds the store itself, and "" for any other. The address composed into every
+// reference the mesh built, at the moment it is used and never before (novox/hq 04-ISSUES/102).
+//
+// **Genesis places the network after the store, the broker, the vault and the catalogue.** Each
+// of those is built and pushed to a node that is on no network, and the store is on that same
+// node; an answer of "no store" there would refuse every one of those pushes and hand every one
+// of those builds a base nothing can pull. Loopback is the truth on that machine, and it is the
+// address genesis itself reaches the store by.
+func artifactStoreAddress(ctx context.Context, inv *inventory.Inventory,
+	shelf map[string]catalogue.Manifest, forNode string) (string, error) {
+	onNetwork, err := whereEveryoneIs(ctx, inv, shelf)
+	if err != nil {
+		return "", err
+	}
+	on := map[string]bool{}
+	for name := range onNetwork {
+		on[name] = true
+	}
+	node, port, found, err := artifactStoreOnNetwork(ctx, inv, on)
+	if err != nil {
+		return "", err
+	}
+	if found {
+		return overlay.InternalName(node) + ":" + port, nil
+	}
+	holder, port, found, err := artifactStoreHolder(ctx, inv)
+	if err != nil || !found || holder != forNode {
+		return "", err
+	}
+	return "127.0.0.1:" + port, nil
+}
+
+// artifactStoreHolder is whichever node is assigned a module offering the artifact store, on or
+// off the network, and the port that node put it on.
+func artifactStoreHolder(ctx context.Context, inv *inventory.Inventory) (node, port string, found bool, err error) {
+	nodes, err := inv.Nodes(ctx)
+	if err != nil {
+		return "", "", false, err
+	}
+	all := map[string]bool{}
+	for _, n := range nodes {
+		all[n.Name] = true
+	}
+	return artifactStoreOnNetwork(ctx, inv, all)
+}

cmd/mesh-controller/network_test.go

@@ -0,0 +1,337 @@
+package main
+
+import (
+	"context"
+	"os"
+	"reflect"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/overlay"
+)
+
+// placementOf is what the mesh holds about where one node is.
+func placementOf(t *testing.T, ctx context.Context, inv *inventory.Inventory, name string) inventory.Overlay {
+	t.Helper()
+	placed, err := inv.Overlays(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, one := range placed {
+		if one.Name == name {
+			return one
+		}
+	}
+	t.Fatalf("%s is not placed at all", name)
+	return inventory.Overlay{}
+}
+
+// The sibling of `node public-domain`, and the worse one: a placement is three facts declared
+// together, so an invocation that said none of them took all three away — the endpoint every other
+// machine dials, the site, and the hub. A mesh whose hub was placed that way has no paths left.
+func TestPlacingANodeWithNothingSaidDoesNotUnplaceIt(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	if err := overlayPlace(ctx, open.inventory,
+		[]string{"anchor", "--endpoint", "198.51.100.10:51820", "--site", "hosting", "--hub"}); err != nil {
+		t.Fatal(err)
+	}
+
+	err := overlayPlace(ctx, open.inventory, []string{"anchor"})
+	if err == nil {
+		t.Fatal("saying nothing unplaced the node instead of being refused")
+	}
+	if !strings.Contains(err.Error(), "--nothing") {
+		t.Errorf("the refusal does not say how to mean it: %v", err)
+	}
+
+	held := placementOf(t, ctx, open.inventory, "anchor")
+	if held.Endpoint != "198.51.100.10:51820" || held.Site != "hosting" || !held.Hub {
+		t.Fatalf("the placement was taken away by an invocation that was refused: %+v", held)
+	}
+}
+
+// Placing a machine with nothing set is a real thing to want — one that roams and opens every path
+// itself is exactly that — so it keeps a way to be said, by name.
+func TestPlacingANodeWithNothingIsAskedForByName(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	if err := overlayPlace(ctx, open.inventory,
+		[]string{"anchor", "--endpoint", "198.51.100.10:51820", "--site", "hosting"}); err != nil {
+		t.Fatal(err)
+	}
+	if err := overlayPlace(ctx, open.inventory, []string{"anchor", "--nothing"}); err != nil {
+		t.Fatal(err)
+	}
+	held := placementOf(t, ctx, open.inventory, "anchor")
+	if held.Endpoint != "" || held.Site != "" || held.Hub {
+		t.Fatalf("--nothing did not place it with nothing: %+v", held)
+	}
+}
+
+// Both at once cannot be meant, so neither silently wins.
+func TestAPlacementAndNothingTogetherIsRefused(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	if err := overlayPlace(ctx, open.inventory,
+		[]string{"anchor", "--site", "hosting", "--nothing"}); err == nil {
+		t.Fatal("a placement and --nothing together was accepted")
+	}
+}
+
+// A machine is named for the others only while it is on the private network — placed AND running
+// what puts it there — the same set the resolver means by "on the private network". A machine
+// that has an address and no networking is not named: a name resolving to an address that does
+// not answer hangs where an unknown name fails at once (novox/hq issue 079).
+func TestOnlyAMachineOnThePrivateNetworkIsNamed(t *testing.T) {
+	open := aMesh(t) // two placed machines, both assigned what puts them on the private network
+	ctx := t.Context()
+	shelf, err := open.inventory.Catalogue(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	names, err := namesInTheMesh(ctx, open.inventory, shelf)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if names["anchor.internal"] != "10.77.0.1" || names["laptop.internal"] != "10.77.0.2" {
+		t.Fatalf("two machines on the network are not both named: %v", names)
+	}
+	// The laptop keeps its place and its address, and stops running the network.
+	if err := open.inventory.Unassign(ctx, "laptop", overlay.Name); err != nil {
+		t.Fatal(err)
+	}
+	names, err = namesInTheMesh(ctx, open.inventory, shelf)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, still := names["laptop.internal"]; still || names["anchor.internal"] != "10.77.0.1" {
+		t.Fatalf("a machine that left the network is still named, or the one that stayed is not: %v", names)
+	}
+}
+
+// novox/hq ADR 0105: the range every address is composed from is the adopted tunnel's, read from
+// the tunnel the hub holds — never stored anywhere else.
+func TestTheOverlaysRangeIsTheAdoptedTunnels(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+	t.Setenv(OverlayCIDRVar, "10.99.0.0/16")
+
+	before, err := overlayRange(ctx, inv)
+	if err != nil || before != "10.99.0.0/16" {
+		t.Fatalf("without an adopted tunnel the range is not what genesis said: %q %v", before, err)
+	}
+
+	// The hub becomes what genesis makes of a machine in use: adopted, enrolled with the found
+	// tunnel's key, and presenting the tunnel.
+	if err := inv.SetAdopted(ctx, "anchor", true); err != nil {
+		t.Fatal(err)
+	}
+	hub, err := inv.NodeByName(ctx, "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	const key = "THE-TUNNELS-KEY========================="
+	if err := inv.RecordOverlayKey(ctx, hub.ID, key); err != nil {
+		t.Fatal(err)
+	}
+	if err := inv.RecordTunnel(ctx, hub.ID, inventory.Tunnel{
+		Interface: "wg0", Unit: "wg-quick@wg0", Config: "/etc/wireguard/wg0.conf", Port: 51900,
+		Address: "192.0.2.1/24", Range: "192.0.2.0/24", PublicKey: key,
+		Peers: []inventory.TunnelPeer{{PublicKey: "PEER-TWO", Address: "192.0.2.2"}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	after, err := overlayRange(ctx, inv)
+	if err != nil || after != "192.0.2.0/24" {
+		t.Fatalf("with an adopted tunnel the range is %q (%v), not the tunnel's", after, err)
+	}
+
+	// A placement whose endpoint is on another port than the tunnel's is refused: the peers dial
+	// the tunnel's port.
+	err = overlayPlace(ctx, inv, []string{"anchor", "--endpoint", "198.51.100.10:51820", "--site", "hosting", "--hub"})
+	if err == nil || !strings.Contains(err.Error(), "51900") {
+		t.Fatalf("an endpoint off the tunnel's port was accepted: %v", err)
+	}
+	// On the tunnel's port, the hub is placed at the tunnel's address — whatever it had before.
+	if err := inv.SetPlace(ctx, "anchor", "", "", false, ""); err != nil {
+		t.Fatal(err)
+	}
+	if err := overlayPlace(ctx, inv, []string{"anchor", "--endpoint", "198.51.100.10:51900", "--site", "hosting", "--hub"}); err != nil {
+		t.Fatal(err)
+	}
+	if placed := placementOf(t, ctx, inv, "anchor"); placed.Address != "192.0.2.1" {
+		t.Fatalf("the hub was placed at %s, not the tunnel's own address", placed.Address)
+	}
+
+	// And the hub's declaration carries the peer and the takeover.
+	nodes, computed, err := graph(ctx, open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var hubNode overlay.Node
+	for _, n := range nodes {
+		if n.Name == "anchor" {
+			hubNode = n
+		}
+	}
+	if hubNode.TakesOver == nil || hubNode.TakesOver.Unit != "wg-quick@wg0" {
+		t.Errorf("the hub is not told to take over the found tunnel: %+v", hubNode)
+	}
+	carried := false
+	for _, p := range computed["anchor"] {
+		if p.Key == "PEER-TWO" && p.Allowed == "192.0.2.2/32" {
+			carried = true
+		}
+	}
+	if !carried {
+		t.Errorf("the hub's peer list does not carry the tunnel's peer: %+v", computed["anchor"])
+	}
+}
+
+// A takeover is composed only for a hub whose placement agrees with the tunnel: an address or an
+// endpoint port that differs would have the host stop the found interface and raise the mesh's
+// where no peer is listening.
+func TestATakeoverIsNotComposedForAHubPlacedOffItsTunnel(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+	if err := inv.SetAdopted(ctx, "anchor", true); err != nil {
+		t.Fatal(err)
+	}
+	hub, err := inv.NodeByName(ctx, "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	const key = "THE-TUNNELS-KEY========================="
+	if err := inv.RecordOverlayKey(ctx, hub.ID, key); err != nil {
+		t.Fatal(err)
+	}
+	if err := inv.RecordTunnel(ctx, hub.ID, inventory.Tunnel{
+		Interface: "wg0", Unit: "wg-quick@wg0", Config: "/etc/wireguard/wg0.conf", Port: 51900,
+		Address: "192.0.2.1/24", Range: "192.0.2.0/24", PublicKey: key}); err != nil {
+		t.Fatal(err)
+	}
+	// aMesh placed anchor at 10.77.0.1 on :51820 — the record of a hub placed before it took the
+	// tunnel over.
+	_, _, err = graph(ctx, open)
+	if err == nil {
+		t.Fatal("a takeover was composed for a hub whose address and port are not the tunnel's")
+	}
+	for _, want := range []string{"10.77.0.1", "192.0.2.1", "51820", "51900", "overlay place anchor"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Errorf("the refusal does not say %q: %v", want, err)
+		}
+	}
+	// Re-placed on the tunnel, it composes.
+	if err := inv.SetPlace(ctx, "anchor", "", "", false, ""); err != nil {
+		t.Fatal(err)
+	}
+	if err := overlayPlace(ctx, inv, []string{"anchor", "--endpoint", "198.51.100.10:51900", "--site", "here", "--hub"}); err != nil {
+		t.Fatal(err)
+	}
+	if _, _, err := graph(ctx, open); err != nil {
+		t.Fatalf("re-placed on the tunnel, the graph still refuses: %v", err)
+	}
+}
+
+// theResolver is the catalogue's dnsmasq module as it is, or the test is skipped where the
+// catalogue is not beside this checkout.
+func theResolver(t *testing.T) catalogue.Manifest {
+	t.Helper()
+	raw, err := os.ReadFile("../../../mesh-catalog/modules/dnsmasq/module.json")
+	if err != nil {
+		t.Skipf("the catalogue is not beside this checkout: %v", err)
+	}
+	m, err := catalogue.ParseManifest(raw)
+	if err != nil {
+		t.Fatalf("dnsmasq does not parse:\n%v", err)
+	}
+	return m
+}
+
+// The resolver is handed every machine on the private network as a wildcard, and is handed it again
+// when a machine leaves — through the module's own manifest asking for the fact, with no module of the
+// mesh's own in between (hal dnsmasq-app conversion, novox/hq 08-connectivity). It is the mesh's one
+// resolver (ADR 0194), and the container runtime is given no resolver of its own (ADR 0196).
+func TestTheResolverIsToldEveryMachineOnTheNetworkAndToldAgainWhenOneLeaves(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, theResolver(t))
+	if _, err := assign(ctx, open, "anchor", "dnsmasq"); err != nil {
+		t.Fatal(err)
+	}
+	// Its bus credential, as assigning issues it where the bus is reachable (novox/hq issue 203):
+	// no bus is known to this test, so it is minted here, or composing refuses the placeholder.
+	if _, err := open.inventory.MintBusPassword(ctx, inventory.BusUser{
+		Username: "anchor.dnsmasq", Kind: inventory.BusModule, Node: "anchor", Module: "dnsmasq"}); err != nil {
+		t.Fatal(err)
+	}
+	zones := func() string {
+		t.Helper()
+		for _, r := range composed(t, open, "anchor").Resources {
+			if r["id"] == "dnsmasq.fact-node-zones" {
+				if r["path"] != "/etc/mesh-resolver/nodes.conf" {
+					t.Fatalf("the machines were written somewhere the resolver does not read: %v", r["path"])
+				}
+				return r["content"].(string)
+			}
+		}
+		t.Fatal("the resolver was not handed the machines")
+		return ""
+	}
+	first := zones()
+	for _, want := range []string{
+		"local=/internal/", "address=/anchor.internal/10.77.0.1\n", "address=/laptop.internal/10.77.0.2\n",
+	} {
+		if !strings.Contains(first, want) {
+			t.Errorf("the resolver's machines lack %q:\n%s", want, first)
+		}
+	}
+	// The container runtime is given no resolver of its own (novox/hq ADR 0196): it copies its
+	// machine's, which name the mesh's resolver first. A `dns` key would be a second account of where a
+	// container asks, read only when the runtime starts.
+	for _, r := range composed(t, open, "anchor").Resources {
+		if r["id"] == "dnsmasq.runtime-dns" {
+			t.Errorf("the resolver still writes the runtime's own dns: %v", r)
+		}
+	}
+
+	// The laptop keeps its place and its address, and stops running the network.
+	if err := open.inventory.Unassign(ctx, "laptop", overlay.Name); err != nil {
+		t.Fatal(err)
+	}
+	after := zones()
+	if strings.Contains(after, "laptop") || !strings.Contains(after, "address=/anchor.internal/10.77.0.1\n") {
+		t.Fatalf("a machine that left the network is still a wildcard, or the one that stayed is not:\n%s", after)
+	}
+}
+
+// The roster is the machines and nothing else (novox/hq ADR 0191): each node's internal domain covers
+// every route on it, and a node's public domains are public DNS's. A routed name in `.Names` was a
+// private answer for a public name, handed by a resolver serving a LAN to a phone that could not use it.
+func TestTheRosterNamesOnlyTheMachines(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	gens, err := generators(ctx, open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, node := range []string{"anchor", "laptop"} {
+		plan, settings, err := planFor(ctx, open, node)
+		if err != nil {
+			t.Fatal(err)
+		}
+		with, _, err := renderingFor(ctx, open, node, plan, settings, gens, Reading)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if !reflect.DeepEqual(with.Names, with.Machines) {
+			t.Fatalf("%s's roster names more than the machines:\n names    %v\n machines %v", node, with.Names, with.Machines)
+		}
+	}
+}

cmd/mesh-controller/nodes.go

@@ -0,0 +1,573 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/token"
+)
+
+// what a machine is, and what it is allowed to be told.
+//
+// Split out of main.go, which had reached 2,769 lines because appending was always the
+// cheapest next step. That is how novox/hq ADR 0001 records `hal/sdk` reaching 34,636:
+// nothing in it was wrong, and no one edit was the one that should have been a new file.
+
+func nodeCommand(ctx context.Context, args []string) error {
+	if len(args) == 0 {
+		return errors.New("node add <name>, node list, node show <name>, or " + publicDomainUsage)
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	switch args[0] {
+	case "show":
+		if len(args) != 2 {
+			return errors.New("node show <name>")
+		}
+		return showNode(ctx, inv, args[1])
+	case "add":
+		return addNode(ctx, inv, args[1:])
+
+	case "list":
+		nodes, err := inv.Nodes(ctx)
+		if err != nil {
+			return err
+		}
+		// **The same list, for something other than a person** — the console's discovery reads it
+		// (novox/hq ADR 0195), and a reader that parses a printed column breaks when it is reworded.
+		if len(args) > 1 && args[1] == "--json" {
+			type listed struct {
+				Name  string `json:"name"`
+				Heard string `json:"heard"`
+				Mode  string `json:"mode"`
+				ID    string `json:"id"`
+			}
+			out := make([]listed, 0, len(nodes))
+			for _, n := range nodes {
+				out = append(out, listed{Name: n.Name, Heard: heardFrom(n), Mode: modeOf(n), ID: n.ID})
+			}
+			return printJSON(out)
+		}
+		if len(nodes) == 0 {
+			// Said rather than printed as nothing: an empty list and a failed read must never
+			// look the same, and this command answering "none" is only honest because getting
+			// here means the store answered.
+			fmt.Println("this mesh has no node records yet")
+			return nil
+		}
+		for _, n := range nodes {
+			fmt.Printf("%-20s %-14s %-9s %s\n", n.Name, heardFrom(n), modeOf(n), n.ID)
+		}
+		return nil
+
+	case "public-domain":
+		// The domain this node composes its routed names under (novox/hq ADR 0066).
+		//
+		// **The form with no argument reports; clearing is asked for by name.** It used to clear —
+		// so `node public-domain anchor`, which reads like a question and is what anybody types to
+		// find out what the answer is, silently took every routed name the node had. A read-shaped
+		// invocation must never be a destructive write: there is no output that makes up for it,
+		// because the damage is already done by the time it prints.
+		return publicDomain(ctx, inv, args[1:])
+
+	case "networks":
+		// Removed by novox/hq ADR 0140, which superseded the record that added it. The filter no
+		// longer names any network: it constrains what arrives from outside the machine and says
+		// nothing about what did not, so there is no list to keep. Answered rather than met with
+		// "unknown command", because this was the documented way to stop a flip cutting a machine's
+		// containers off and somebody will reasonably still type it.
+		return errors.New("`node networks` is gone (novox/hq ADR 0140). The filter constrains what " +
+			"arrives from outside this machine and says nothing about traffic that did not, so no " +
+			"network is named anywhere and nothing needs to be said to keep a machine's own " +
+			"containers reaching outward. The machine reports which of its links face outside; see " +
+			"`node show <name>`")
+
+	case "account":
+		// The operator's login on this machine (novox/hq to-be 29): what a home-scoped file is
+		// owned by and which account `ssh <node>` uses. Reports with no argument; sets with one;
+		// an optional second argument is the home when it is not /home/<account>.
+		return nodeAccount(ctx, inv, args[1:])
+
+	default:
+		return fmt.Errorf("node has no %q; it has add, list, show, public-domain and account", args[0])
+	}
+}
+
+// addNode creates a node record, adopted when the operator says so (novox/hq ADR 0100).
+func addNode(ctx context.Context, inv *inventory.Inventory, args []string) error {
+	set := flag.NewFlagSet("node add", flag.ContinueOnError)
+	adopted := set.Bool("adopted", false,
+		"the machine is in use: keep what is found on it until each module is taken")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 1 {
+		return errors.New("node add <name> [--adopted]")
+	}
+	node, err := inv.AddNodeAs(ctx, positionals[0], *adopted)
+	if err != nil {
+		return err
+	}
+	fmt.Printf("added %s (%s)", node.Name, node.ID)
+	if node.Adopted {
+		fmt.Print(", adopted")
+	}
+	fmt.Println()
+	return nil
+}
+
+// modeOf is a node's mode as a word (novox/hq ADR 0100): an adopted node is said to be adopted
+// wherever the mesh reports a node's state.
+func modeOf(n inventory.Node) string {
+	if n.Adopted {
+		return "adopted"
+	}
+	return "converged"
+}
+
+// publicDomainUsage is the one description of the three forms, so a refusal and the help agree.
+// nodeAccount reports or sets a node's operator account (novox/hq to-be 29). Read-shaped with no
+// argument, like public-domain: `node account novox` answers, it does not change anything.
+func nodeAccount(ctx context.Context, inv *inventory.Inventory, positionals []string) error {
+	if len(positionals) == 0 || len(positionals) > 3 {
+		return errors.New("node account <name> — what it is now; " +
+			"node account <name> <account> [home] — set it (home defaults to /home/<account>)")
+	}
+	node := positionals[0]
+	if len(positionals) == 1 {
+		who, err := inv.NodeByName(ctx, node)
+		if err != nil {
+			return err
+		}
+		if who.Account == "" {
+			fmt.Printf("%s has no operator account known\n", node)
+			fmt.Printf("  `node account %s <account>` sets it\n", node)
+			return nil
+		}
+		fmt.Printf("%s logs a person in as %s (home %s)\n", node, who.Account, who.Home())
+		return nil
+	}
+	home := ""
+	if len(positionals) == 3 {
+		home = positionals[2]
+	}
+	if err := inv.SetAccount(ctx, node, positionals[1], home); err != nil {
+		return err
+	}
+	fmt.Printf("%s logs a person in as %s\n", node, positionals[1])
+	fmt.Printf("  run `push %s` once ssh-client is assigned, to send its operator config\n", node)
+	return nil
+}
+
+const publicDomainUsage = "node public-domain <name> — what it is now; " +
+	"<name> <domain> to set it; <name> --clear to take it away"
+
+// publicDomain reads, sets or clears the domain a node composes its routed names under.
+//
+// Three forms, and the destructive one is the only one that has to be asked for. Clearing is a
+// real thing to want — a machine that stops facing the outside composes no names, and
+// lab-versus-production is this one setting (novox/hq ADR 0066) — so it keeps a way to say it.
+// What it does not keep is being the thing that happens when nothing was said at all.
+func publicDomain(ctx context.Context, inv *inventory.Inventory, args []string) error {
+	set := flag.NewFlagSet("node public-domain", flag.ContinueOnError)
+	clear := set.Bool("clear", false, "take the domain away; it composes no routed names after")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) == 0 || len(positionals) > 2 {
+		return errors.New(publicDomainUsage)
+	}
+	node := positionals[0]
+
+	switch {
+	case *clear && len(positionals) == 2:
+		// Both, which cannot be meant. Refused rather than one of them silently winning.
+		return fmt.Errorf("give %s a domain or --clear, not both: %q and --clear say opposite "+
+			"things and the mesh will not choose between them", node, positionals[1])
+
+	case *clear:
+		if err := inv.SetPublicDomain(ctx, node, ""); err != nil {
+			return err
+		}
+		fmt.Printf("%s has no public domain, so it composes no routed names\n", node)
+		fmt.Printf("  run `push %s` to take them off it\n", node)
+		return nil
+
+	case len(positionals) == 2:
+		if err := inv.SetPublicDomain(ctx, node, positionals[1]); err != nil {
+			return err
+		}
+		fmt.Printf("%s composes its routed names under %s\n", node, positionals[1])
+		fmt.Printf("  run `push %s` to send it\n", node)
+		return nil
+
+	default:
+		// Asked, so answered. NodeByName first, so a name the mesh has never heard of is a refusal
+		// rather than "it has no public domain", which is true of that name and says nothing.
+		if _, err := inv.NodeByName(ctx, node); err != nil {
+			return err
+		}
+		domain, err := inv.PublicDomainOf(ctx, node)
+		if err != nil {
+			return err
+		}
+		if domain == "" {
+			fmt.Printf("%s has no public domain, so it composes no routed names\n", node)
+			fmt.Printf("  `node public-domain %s <domain>` gives it one\n", node)
+			return nil
+		}
+		fmt.Printf("%s composes its routed names under %s\n", node, domain)
+		return nil
+	}
+}
+
+func tokenCommand(ctx context.Context, args []string) error {
+	if len(args) == 0 || args[0] != "issue" {
+		return errors.New("token issue --node <name>, or token issue --new <name>")
+	}
+
+	set := flag.NewFlagSet("token issue", flag.ContinueOnError)
+	existing := set.String("node", "", "issue for a node record that already exists")
+	fresh := set.String("new", "", "create the node record, then issue for it")
+	validFor := set.Duration("for", time.Hour, "how long the token may be used")
+	adopted := set.Bool("adopted", false,
+		"the machine joining is in use: it is adopted, and keeps what is found on it")
+	if err := set.Parse(args[1:]); err != nil {
+		return err
+	}
+
+	// Exactly one, because the difference is what the token binds to. A command that guessed
+	// would sometimes create a second record for a machine that already has one.
+	if (*existing == "") == (*fresh == "") {
+		return errors.New("give exactly one of --node <name> or --new <name>: the first is a " +
+			"machine the mesh already has a record for, the second is one it has never seen")
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	issued, err := issueFor(ctx, inv, *existing, *fresh, *adopted, *validFor)
+	if err != nil {
+		return err
+	}
+
+	// Assembled from two contexts by the process that holds both grants. Neither reads the
+	// other's store (novox/hq ADR 0008) — each is asked for its own part.
+	ident, err := openIdentity(ctx)
+	if err != nil {
+		return err
+	}
+	defer ident.Close()
+	key, err := ident.Establish(ctx)
+	if err != nil {
+		return err
+	}
+
+	// The account is created before the token is handed over, which is what removes the
+	// chicken-and-egg entirely: the mesh runs the broker, so a joining node's credentials can
+	// exist before it does. The one-time secret IS the password, so a node's first connection is
+	// already authenticated and enrolment is what happens over it.
+
+	made := token.Token{Node: issued.Node.Name, Signer: key.Public, Secret: issued.Secret,
+		Adopted: issued.Node.Adopted}
+
+	// Absent is a state, not a failure: a control plane can hold records and a key before it has
+	// a broker. What it cannot do is issue a token anybody could use, and Missing() says so.
+	known, err := broker.FromEnvironment()
+	switch {
+	case err == nil:
+		made.Broker, made.Fingerprint = known.Address, known.Fingerprint
+	case errors.Is(err, broker.ErrNotConfigured):
+	default:
+		return err
+	}
+	encoded, err := made.Encode()
+	if err != nil {
+		return err
+	}
+
+	joins := ""
+	if made.Adopted {
+		joins = ", joining adopted"
+	}
+	fmt.Printf("token for %s%s, usable once, until %s\n\n  %s\n\n",
+		issued.Node.Name, joins, issued.Expires.Format(time.RFC3339), encoded)
+	fmt.Println("This is the only time it is shown. What is stored is a hash of the secret.")
+
+	if missing := made.Missing(); len(missing) > 0 {
+		fmt.Printf("\nINCOMPLETE — this token cannot be used to join anything yet. Missing:\n")
+		for _, m := range missing {
+			fmt.Printf("  - %s\n", m)
+		}
+		fmt.Printf("\nSet %s and %s once the broker is raised.\n",
+			broker.AddressVar, broker.CertificateVar)
+	}
+	return nil
+}
+
+// issueFor is the inventory's half of issuing a token: the record, made when it is new, adopted
+// when the operator says so, and the one-time secret for it. The node in what it returns carries
+// its mode, which is what the token says.
+func issueFor(ctx context.Context, inv *inventory.Inventory, existing, fresh string, adopted bool,
+	validFor time.Duration) (inventory.Issued, error) {
+	name := existing
+	if fresh != "" {
+		node, err := inv.AddNodeAs(ctx, fresh, adopted)
+		if err != nil {
+			return inventory.Issued{}, err
+		}
+		name = node.Name
+	}
+	// Not saying adopted leaves the node as it is: re-issuing a token for an adopted node does not
+	// converge it — converging is its own act, previewed (novox/hq ADR 0100). And saying it for a
+	// node already converged is refused rather than done quietly: returning a node to adopted is
+	// its own act too, which unloads the mesh's filter and enables the found firewall again.
+	if adopted && fresh == "" {
+		node, err := inv.NodeByName(ctx, name)
+		if err != nil {
+			return inventory.Issued{}, err
+		}
+		if !node.Adopted {
+			return inventory.Issued{}, fmt.Errorf("%s is converged, and a token does not change "+
+				"that: run `adopt %s` to return it to adopted, then issue the token without "+
+				"--adopted", name, name)
+		}
+	}
+
+	return inv.IssueToken(ctx, name, validFor)
+}
+
+func identityCommand(ctx context.Context, args []string) error {
+	if len(args) == 0 || args[0] != "show" {
+		return errors.New("identity show")
+	}
+	ident, err := openIdentity(ctx)
+	if err != nil {
+		return err
+	}
+	defer ident.Close()
+
+	// Establish rather than read: a control plane asked for its identity before it has one should
+	// get one, not an error. Generating it is idempotent, so this is safe to run at any time.
+	key, err := ident.Establish(ctx)
+	if err != nil {
+		return err
+	}
+	fmt.Printf("signing key   %s\n", key.ID)
+	fmt.Printf("fingerprint   %s\n", key.Fingerprint())
+	fmt.Printf("created       %s\n", key.Created.Format(time.RFC3339))
+	fmt.Printf("\nThe public half of this travels in every enrolment token. A node believes a\n" +
+		"declaration because it carries a signature this key made (novox/hq ADR 0004).\n")
+	return nil
+}
+
+func brokerCommand(ctx context.Context, args []string) error {
+	if len(args) > 0 && args[0] == "certificate" {
+		return busCertificate(args[1:])
+	}
+	if len(args) > 0 && args[0] == "accounts" {
+		return busAccounts(ctx, args[1:])
+	}
+	if len(args) > 0 && args[0] == "consumer-reset" {
+		return consumerReset(args[1:])
+	}
+	if len(args) == 0 || args[0] != "show" {
+		return errors.New("broker show | broker certificate [--check] --into <directory> | broker accounts --into <file> | " +
+			"broker consumer-reset <stream> <consumer>")
+	}
+	known, err := broker.FromEnvironment()
+	if errors.Is(err, broker.ErrNotConfigured) {
+		fmt.Printf("no broker configured. Set %s and %s.\n\n"+
+			"Until then tokens carry the signing key and the one-time secret, and say what they\n"+
+			"are missing. They cannot be used to join.\n",
+			broker.AddressVar, broker.CertificateVar)
+		return nil
+	}
+	if err != nil {
+		return err
+	}
+	fmt.Printf("address       %s\n", known.Address)
+	fmt.Printf("fingerprint   %s\n", known.Fingerprint)
+	fmt.Print("\nThe fingerprint is computed from the certificate on disk, never configured. A\n" +
+		"node checks it before sending anything (novox/hq ADR 0004).\n")
+	return nil
+}
+
+// consumerReset re-makes one consumer on a stream that keeps history to start from now (novox/hq issue
+// 248): the way out of a consumer replaying a week of announcements, said rather than done by hand. A
+// person's act — what was pending is dropped — so it is a command, and nothing calls it on its own.
+func consumerReset(args []string) error {
+	if len(args) != 2 {
+		return errors.New("broker consumer-reset <stream> <consumer>, e.g. broker consumer-reset EVENTS controller")
+	}
+	address, err := broker.BusAddress()
+	if err != nil {
+		return err
+	}
+	js, err := broker.Dial(address)
+	if err != nil {
+		return fmt.Errorf("cannot reach the bus: %w", err)
+	}
+	defer js.Close()
+	before, after, err := js.ResetConsumer(args[0], args[1])
+	if err != nil {
+		return err
+	}
+	fmt.Printf("consumer %s on %s re-made to deliver from now\n", args[1], args[0])
+	fmt.Printf("  before: delivers %s, delivered to %d, acknowledged to %d, %d pending, %d unacknowledged\n",
+		before.DeliverPolicy, before.Delivered, before.AckFloor, before.Pending, before.AckPending)
+	fmt.Printf("  after:  delivers %s, %d pending; what was pending is dropped. A holder bound to it may need its "+
+		"process restarted to bind again\n", after.DeliverPolicy, after.Pending)
+	return nil
+}
+
+// heardFrom says when a node was last heard from, in a form somebody can act on.
+//
+// "never" and "an hour ago" are different answers and are kept different. A node that has never
+// spoken did not finish joining; a node last heard from an hour ago is running an hour-old
+// picture of the mesh.
+func heardFrom(n inventory.Node) string {
+	silent, ever := n.Silent()
+	switch {
+	case !ever:
+		return "never spoken"
+	case silent > SilentFor:
+		return "out of touch " + roughly(silent)
+	default:
+		return "here"
+	}
+}
+
+// roughly is a duration a person reads rather than parses.
+func roughly(d time.Duration) string {
+	switch {
+	case d < time.Hour:
+		return fmt.Sprintf("%dm", int(d.Minutes()))
+	case d < 48*time.Hour:
+		return fmt.Sprintf("%dh", int(d.Hours()))
+	default:
+		return fmt.Sprintf("%dd", int(d.Hours()/24))
+	}
+}
+
+// showNode says what one machine reported about itself, in its own words.
+//
+// **A capability is detected and never assumed** (novox/hq ADR 0009), so the only account of what
+// a machine can do is the one it gave — and its detail is half of that account. The mesh was
+// keeping the yes and discarding the reason, which makes *this machine has no seat* an answer with
+// nowhere to go: a person told a machine lacks something wants to know what the detector saw.
+//
+// It is also where "what should it be configured as" is read. The same line that gates an
+// assignment carries `card1-DP-1`, and a person composing settings for that machine needs it.
+func showNode(ctx context.Context, inv *inventory.Inventory, name string) error {
+	node, err := inv.NodeByName(ctx, name)
+	if err != nil {
+		return err
+	}
+	fmt.Printf("%s\n", node.Name)
+	fmt.Printf("  last heard from   %s\n", heardFrom(node))
+	// Which host runs it, as it reported (novox/hq 04-ISSUES/087). Said whenever known, because a
+	// host refuses a declaration carrying a field it does not understand and refuses it WHOLE — so
+	// which host a machine runs is what decides whether the mesh can send it anything new, and
+	// nothing could say it. "not reported" rather than blank: a machine that has not said is a
+	// different thing from one running nothing.
+	fmt.Printf("  host              %s\n", orNotReported(node.HostVersion))
+	if err := showMode(ctx, inv, node); err != nil {
+		return err
+	}
+
+	// The domain its routed names are composed under, when it has one (novox/hq ADR 0066). Shown
+	// only when set: a machine that serves nothing to the outside has no domain, and saying so of
+	// every internal node would be noise.
+	domain, err := inv.PublicDomainOf(ctx, name)
+	if err != nil {
+		return err
+	}
+	if domain != "" {
+		fmt.Printf("  public domain     %s\n", domain)
+	}
+
+	// A provider here failing a consumer, or a consumer here failed (novox/hq ADR 0224). Before the
+	// capabilities, because it is something not working now and they are a description.
+	failing, err := failingProviders(ctx, inv)
+	if err != nil {
+		return err
+	}
+	if here := failingOn(failing, name); len(here) > 0 {
+		fmt.Printf("\n  %d consumer(s) a provider keeps failing, here or for a module here:\n", len(here))
+		for _, line := range failingLines(here, time.Now()) {
+			fmt.Printf("  %s\n", line)
+		}
+	}
+
+	held, err := inv.Profile(ctx, name)
+	if err != nil {
+		return err
+	}
+	if held == nil {
+		// Never reported is not the same as reported nothing, and the remedy differs: one is a
+		// machine that has not run the host yet, the other is a machine that ran it and can do
+		// nothing.
+		fmt.Printf("\n  this machine has never said what it can do, so everything requiring a\n" +
+			"  capability is refused here — run the host on it\n")
+		return nil
+	}
+	if len(held) == 0 {
+		fmt.Printf("\n  it reported no capabilities at all\n")
+		return nil
+	}
+
+	fmt.Printf("\n  what it can do, as it reported:\n")
+	for _, c := range held {
+		mark := "no "
+		if c.Present {
+			mark = "yes"
+		}
+		fmt.Printf("    %s  %-20s %s\n", mark, c.Name, c.Detail)
+	}
+
+	assigned, err := inv.Assigned(ctx, name)
+	if err != nil {
+		return err
+	}
+	if len(assigned) > 0 {
+		fmt.Printf("\n  assigned: %s\n", strings.Join(assigned, ", "))
+	}
+	return nil
+}
+
+// orNotReported is a fact a machine states about itself, or the fact that it has not.
+func orNotReported(s string) string {
+	if strings.TrimSpace(s) == "" {
+		return "not reported — this machine has not said since the mesh began keeping it"
+	}
+	return s
+}
+
+// printJSON prints a value as indented JSON, the shape every `--json` answers in.
+func printJSON(v any) error {
+	body, err := json.MarshalIndent(v, "", "  ")
+	if err != nil {
+		return err
+	}
+	fmt.Println(string(body))
+	return nil
+}

cmd/mesh-controller/nodes_test.go

@@ -0,0 +1,154 @@
+package main
+
+import (
+	"strings"
+	"testing"
+	"time"
+)
+
+// **A read-shaped invocation is never a destructive write.**
+//
+// `node public-domain anchor` used to clear the domain. It reads like a question — it is what
+// anybody types to find out what the answer is — and it silently took every routed name the node
+// had. There is no output that makes up for that: by the time it prints, the fact is gone.
+func TestAskingForANodesPublicDomainDoesNotTakeItAway(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	if err := open.inventory.SetPublicDomain(ctx, "anchor", "example.test"); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := publicDomain(ctx, open.inventory, []string{"anchor"}); err != nil {
+		t.Fatal(err)
+	}
+	domain, err := open.inventory.PublicDomainOf(ctx, "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if domain != "example.test" {
+		t.Fatalf("asking what the domain is took it away: %q", domain)
+	}
+}
+
+// Clearing is a real thing to want — a machine that stops facing the outside composes no names —
+// so it keeps a way to be said. What it stops being is what happens when nothing was said.
+func TestClearingANodesPublicDomainIsAskedForByName(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	if err := open.inventory.SetPublicDomain(ctx, "anchor", "example.test"); err != nil {
+		t.Fatal(err)
+	}
+	if err := publicDomain(ctx, open.inventory, []string{"anchor", "--clear"}); err != nil {
+		t.Fatal(err)
+	}
+	domain, err := open.inventory.PublicDomainOf(ctx, "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if domain != "" {
+		t.Fatalf("--clear did not clear it: %q", domain)
+	}
+}
+
+// A domain sets it, as it always did.
+func TestGivingANodeAPublicDomainSetsIt(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	if err := publicDomain(ctx, open.inventory, []string{"anchor", "example.test"}); err != nil {
+		t.Fatal(err)
+	}
+	domain, err := open.inventory.PublicDomainOf(ctx, "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if domain != "example.test" {
+		t.Fatalf("got %q", domain)
+	}
+}
+
+// A domain and --clear say opposite things. Refused rather than one of them silently winning.
+func TestADomainAndClearTogetherIsRefused(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	if err := open.inventory.SetPublicDomain(ctx, "anchor", "example.test"); err != nil {
+		t.Fatal(err)
+	}
+	err := publicDomain(ctx, open.inventory, []string{"anchor", "other.test", "--clear"})
+	if err == nil {
+		t.Fatal("a domain and --clear together were accepted")
+	}
+	if !strings.Contains(err.Error(), "not both") {
+		t.Fatalf("the refusal does not say why: %v", err)
+	}
+	// And neither half happened.
+	domain, err := open.inventory.PublicDomainOf(ctx, "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if domain != "example.test" {
+		t.Fatalf("a refused command changed something: %q", domain)
+	}
+}
+
+// Asking about a name the mesh has never heard of is a refusal, not "it has no public domain" —
+// which is true of that name and says nothing.
+func TestAskingAboutAMachineTheMeshHasNeverHeardOfIsRefused(t *testing.T) {
+	open := aMesh(t)
+	if err := publicDomain(t.Context(), open.inventory, []string{"nowhere"}); err == nil {
+		t.Fatal("a name the mesh does not know was answered as if it were a machine")
+	}
+}
+
+// novox/hq ADR 0100: the operator says a node is adopted — `node add --adopted` for a record, and
+// the token for a machine joining.
+func TestANodeAddedAdoptedIsAdopted(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	if err := addNode(ctx, open.inventory, []string{"joiner", "--adopted"}); err != nil {
+		t.Fatal(err)
+	}
+	n, err := open.inventory.NodeByName(ctx, "joiner")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !n.Adopted {
+		t.Fatal("node add --adopted made a converged node")
+	}
+	if err := addNode(ctx, open.inventory, []string{"plain"}); err != nil {
+		t.Fatal(err)
+	}
+	if n, _ := open.inventory.NodeByName(ctx, "plain"); n.Adopted {
+		t.Fatal("node add without --adopted made an adopted node")
+	}
+}
+
+func TestATokenIssuedAdoptedSaysSoAndReissuingDoesNotConverge(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	issued, err := issueFor(ctx, open.inventory, "", "joiner", true, time.Hour)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !issued.Node.Adopted {
+		t.Fatal("a token issued --adopted is for a node that is not adopted")
+	}
+	again, err := issueFor(ctx, open.inventory, "joiner", "", false, time.Hour)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !again.Node.Adopted {
+		t.Fatal("re-issuing without --adopted converged the node; converging is its own act")
+	}
+	// --adopted for a node already converged is refused, and points at the act that does it.
+	if _, err := issueFor(ctx, open.inventory, "laptop", "", true, time.Hour); err == nil ||
+		!strings.Contains(err.Error(), "adopt laptop") {
+		t.Fatalf("--adopted on a converged node was not refused: %v", err)
+	}
+	if n, _ := open.inventory.NodeByName(ctx, "laptop"); n.Adopted {
+		t.Fatal("a refused token flipped the node to adopted")
+	}
+	// And said for a node that is adopted already, it is the ordinary re-issue.
+	if _, err := issueFor(ctx, open.inventory, "joiner", "", true, time.Hour); err != nil {
+		t.Fatal(err)
+	}
+}

cmd/mesh-controller/operator.go

@@ -0,0 +1,307 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"os"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/secrets"
+)
+
+// operatorCommand is the mesh's one holder of secrets that is not a machine.
+//
+// **Every secret a module holds for itself is sealed to the node that uses it, and a node whose key
+// is gone takes its secrets with it** — the store's superuser and the broker's administrator among
+// them. novox/hq ADR 0085 (amended) gives them a second recipient: a person, with a sealing key
+// whose private half is made where the operator is and never enters the mesh. Making the key and
+// telling the mesh about it are two commands, on purpose: the first needs no mesh at all and runs
+// wherever the operator keeps things; the second gives the mesh the public half and nothing else.
+// The controller's own container is a scratch image with no writable path, which is the right
+// shape for a program that must hold no key — so the private half could not be written there
+// even by mistake.
+//
+//	operator key make [--out <file>]    make a keypair: private half to the file, public half printed
+//	operator key set <public>           tell the mesh which key to seal to
+//	operator key show                   the public key, its fingerprint, and what it can recover
+const operatorUsage = "operator key make [--out <file>] | operator key set <public> [--replace] | " +
+	"operator key show | operator issue <name> --invokes <tool,tool|*> | operator revoke <name> | " +
+	"operator list"
+
+func operatorCommand(ctx context.Context, args []string) error {
+	if len(args) == 0 {
+		return errors.New(operatorUsage)
+	}
+	// The people who may reach the mesh's tools (design 25 §7). Beside the operator's key because
+	// both answer "who, other than a machine, may do something here" — and a person reading this
+	// command's usage is asking exactly that.
+	switch args[0] {
+	case "issue":
+		return personIssue(ctx, args[1:])
+	case "revoke":
+		return personRevoke(ctx, args[1:])
+	case "list":
+		return personList(ctx)
+	}
+	if len(args) < 2 || args[0] != "key" {
+		return errors.New(operatorUsage)
+	}
+	switch args[1] {
+	case "make":
+		return operatorKeyMake(args[2:])
+	case "set":
+		return operatorKeySet(ctx, args[2:])
+	case "show":
+		return operatorKeyShow(ctx)
+	default:
+		return errors.New(operatorUsage)
+	}
+}
+
+// operatorKeyMake needs no mesh: it is the operator's, run wherever the key will live.
+func operatorKeyMake(args []string) error {
+	set := flag.NewFlagSet("operator key make", flag.ContinueOnError)
+	out := set.String("out", "operator.key",
+		"where to write the private key (0600); keep it off the mesh, and keep it")
+	if err := set.Parse(args); err != nil {
+		return err
+	}
+	public, private, err := secrets.Keypair()
+	if err != nil {
+		return err
+	}
+	// Create-exclusive: a key somebody may still need is never overwritten, and there is no window
+	// between checking and writing in which one could appear.
+	if err := writeNew(*out, []byte(private+"\n")); err != nil {
+		return err
+	}
+	fmt.Printf("operator key %s\n", secrets.Fingerprint(public))
+	fmt.Printf("  private half written to %s (0600) — keep it off the mesh, and keep it\n", *out)
+	fmt.Printf("  public half, to give the mesh with `operator key set`:\n")
+	fmt.Printf("public %s\n", public)
+	return nil
+}
+
+func operatorKeySet(ctx context.Context, args []string) error {
+	rest, flags := split(args)
+	set := flag.NewFlagSet("operator key set", flag.ContinueOnError)
+	replace := set.Bool("replace", false,
+		"replace an existing operator key — secrets sealed to the old one stay sealed to it")
+	if err := set.Parse(flags); err != nil {
+		return err
+	}
+	if len(rest) != 1 {
+		return errors.New(operatorUsage)
+	}
+	public := strings.TrimSpace(rest[0])
+	if _, err := secrets.Seal(public, []byte("probe")); err != nil {
+		return fmt.Errorf("that is not a public sealing key: %w", err)
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	if current, err := inv.OperatorKey(ctx); err != nil {
+		return err
+	} else if current != "" && current != public && !*replace {
+		return fmt.Errorf(
+			"the mesh already has an operator key (%s). Pass --replace to change it — "+
+				"secrets sealed to the current key stay sealed to it until each is issued again",
+			secrets.Fingerprint(current))
+	}
+	orphaned, err := inv.SetOperatorKey(ctx, public)
+	if err != nil {
+		return err
+	}
+	fmt.Printf("operator key %s\n", secrets.Fingerprint(public))
+	fmt.Printf("  the mesh holds the public half only and cannot open what it seals to it;\n")
+	fmt.Printf("  from now on every secret a module holds for itself is sealed to it as well.\n")
+	fmt.Printf("  Secrets made before this cannot be — each is recoverable once issued again\n")
+	if orphaned > 0 {
+		fmt.Printf("  %d secret(s) are sealed to the previous key and stay so until issued again\n", orphaned)
+	}
+	return nil
+}
+
+func operatorKeyShow(ctx context.Context) error {
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+	key, err := inv.OperatorKey(ctx)
+	if err != nil {
+		return err
+	}
+	if key == "" {
+		fmt.Println("the mesh has no operator key; `operator key make` then `operator key set` gives it one")
+		return nil
+	}
+	kept, earlier, unrecoverable, err := inv.KeptForOperator(ctx)
+	if err != nil {
+		return err
+	}
+	fmt.Printf("operator key %s\n  %s\n", secrets.Fingerprint(key), key)
+	fmt.Printf("  %d secret(s) recoverable with it\n", len(kept))
+	if len(earlier) > 0 {
+		fmt.Printf("  %d secret(s) sealed to an earlier operator key — recoverable with that key only, until issued again:\n", len(earlier))
+		for _, k := range earlier {
+			fmt.Printf("    %s  %s  %s  (%s)\n", k.Node, k.Module, k.Name, secrets.Fingerprint(k.Key))
+		}
+	}
+	if len(unrecoverable) > 0 {
+		fmt.Printf("  %d secret(s) not recoverable — made before the mesh had an operator key:\n", len(unrecoverable))
+		for _, k := range unrecoverable {
+			fmt.Printf("    %s  %s  %s\n", k.Node, k.Module, k.Name)
+		}
+	}
+	return nil
+}
+
+// readPrivateKey is the operator's key from the file `operator key make` wrote.
+func readPrivateKey(path string) (string, error) {
+	if path == "" {
+		return "", errors.New("--key <file> names the operator's private key, written by `operator key make`")
+	}
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		return "", err
+	}
+	return strings.TrimSpace(string(raw)), nil
+}
+
+// personIssue gives somebody a credential for the mesh's tools, and prints it once.
+//
+// **Printed, not stored.** The mesh keeps a hash and nothing else, so this is the only moment the
+// credential exists anywhere but on the workstation that will use it — the same contract a token has,
+// and for the same reason: a credential recoverable from the mesh's store has the store's blast
+// radius.
+func personIssue(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("operator issue", flag.ContinueOnError)
+	invokes := set.String("invokes", "", "the tools this person may call, comma-separated, or * for every one")
+	// Flags on either side of the name, because the usage this command prints puts them after it —
+	// and the standard parser stops at the first thing that is not a flag, so the order the command
+	// documents was the one order it refused (2026-09-28).
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 1 {
+		return errors.New("operator issue <name> --invokes <tool,tool|*>")
+	}
+	name := positionals[0]
+	if *invokes == "" {
+		return errors.New(
+			"say what this person may call: --invokes mesh-catalog.catalog_tools,gitea.repo_create, " +
+				"or --invokes '*' for an administrator")
+	}
+	var tools []string
+	for _, t := range strings.Split(*invokes, ",") {
+		if t = strings.TrimSpace(t); t != "" {
+			tools = append(tools, t)
+		}
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	if err := inv.RecordPerson(ctx, inventory.Person{Name: name, Invokes: tools}); err != nil {
+		return err
+	}
+	// Refused here rather than at the next composition, where it would stop the whole file being
+	// written for everybody. A name that cannot be part of a subject is one the server would read as
+	// a wider permission than anybody granted.
+	if _, err := broker.PermissionsFor(broker.Principal{
+		Kind: broker.KindPerson, Module: name, Invokes: tools, PasswordHash: "x",
+	}); err != nil {
+		return err
+	}
+
+	user := broker.Principal{Kind: broker.KindPerson, Module: name}.Username()
+	password, err := inv.MintBusPassword(ctx, inventory.BusUser{Username: user, Kind: inventory.BusPerson})
+	if err != nil {
+		return err
+	}
+
+	where, err := broker.FromEnvironment()
+	if err != nil && !errors.Is(err, broker.ErrNotConfigured) {
+		return err
+	}
+	held, err := json.Marshal(struct {
+		URL         string   `json:"url"`
+		Fingerprint string   `json:"fingerprint,omitempty"`
+		User        string   `json:"user"`
+		Password    string   `json:"password"`
+		Person      string   `json:"person"`
+		Invokes     []string `json:"invokes"`
+	}{
+		URL: "nats://" + where.Address, Fingerprint: where.Fingerprint,
+		User: user, Password: password, Person: name, Invokes: tools,
+	})
+	if err != nil {
+		return err
+	}
+
+	fmt.Printf("issued %s, who may call %s\n", name, strings.Join(tools, ", "))
+	fmt.Println("  this is the only time the credential is printed; the mesh keeps a hash")
+	fmt.Println("  it works once the bus has been told, which is the next push to the machine holding mesh-broker")
+	fmt.Println()
+	fmt.Println(string(held))
+	return nil
+}
+
+func personRevoke(ctx context.Context, args []string) error {
+	if len(args) != 1 {
+		return errors.New("operator revoke <name>")
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+
+	if err := open.inventory.ForgetPerson(ctx, args[0]); err != nil {
+		return err
+	}
+	// **Revoked at the next composition, not now.** The bus's users are a file, so a credential stops
+	// working when the file no longer names it. Said plainly, because "revoked" that still works for
+	// another minute is worth knowing about.
+	fmt.Printf("%s is forgotten, and their credential stops working at the next composition — "+
+		"push the machine holding mesh-broker to make it so\n", args[0])
+	return nil
+}
+
+func personList(ctx context.Context) error {
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+
+	people, err := open.inventory.People(ctx)
+	if err != nil {
+		return err
+	}
+	if len(people) == 0 {
+		fmt.Println("nobody but machines reaches this mesh")
+		return nil
+	}
+	for _, p := range people {
+		fmt.Printf("%-20s %s\n", p.Name, strings.Join(p.Invokes, ", "))
+	}
+	return nil
+}

cmd/mesh-controller/order_test.go

@@ -0,0 +1,245 @@
+package main
+
+import (
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// entry is a module as the catalogue holds it: built, so its manifest carries no `build` any more.
+func entry(module string, _ ...string) inventory.Entry {
+	return inventory.Entry{Manifest: catalogue.Manifest{Module: module}}
+}
+
+// stoodOn is what each module's newest build recorded it was handed.
+func stoodOn(edges map[string][]string) map[string][]string {
+	out := map[string][]string{}
+	for module, bases := range edges {
+		for _, b := range bases {
+			out[module] = append(out[module], catalogue.ArtifactStoreScheme+b+"/runtime@sha256:"+strings.Repeat("0", 64))
+		}
+	}
+	return out
+}
+
+// A module built before the module it stands on is built against the old one and reports success
+// (novox/hq 04-ISSUES/131). So bases come first, however the set arrived.
+func TestBasesAreBuiltBeforeWhatStandsOnThem(t *testing.T) {
+	in := []inventory.Entry{entry("app"), entry("runtime"), entry("other"), entry("base")}
+	edges := stoodOn(map[string][]string{"app": {"runtime"}, "runtime": {"base"}})
+	got := orderByBases(in, edges)
+	pos := map[string]int{}
+	for i, e := range got {
+		pos[e.Manifest.Module] = i
+	}
+	if !(pos["base"] < pos["runtime"] && pos["runtime"] < pos["app"]) {
+		t.Fatalf("bases not first: %v", pos)
+	}
+	if len(got) != 4 {
+		t.Fatalf("an entry was lost or doubled: %d", len(got))
+	}
+	// A base outside the set is not waited for: it is not being rebuilt.
+	got = orderByBases([]inventory.Entry{entry("app")}, stoodOn(map[string][]string{"app": {"elsewhere"}}))
+	if len(got) != 1 {
+		t.Fatalf("a dependency outside the set changed the set: %v", got)
+	}
+}
+
+// A module registered from its manifest and never built still names its bases there; once built,
+// the recorded edge is what says so. Both are read, and a module never stands on itself.
+func TestWhatStandsOnAModuleIsReadFromItsBuildOrItsManifest(t *testing.T) {
+	built := entry("gitea")
+	edges := stoodOn(map[string][]string{"gitea": {"mesh-tools"}})
+	if !standsOnModule(built, "mesh-tools", edges) {
+		t.Fatal("a recorded edge was not read")
+	}
+	if standsOnModule(built, "gitea", edges) || standsOnModule(built, "postgres", edges) {
+		t.Fatal("an edge was invented")
+	}
+	fresh := inventory.Entry{Manifest: catalogue.Manifest{Module: "plex", Build: &catalogue.Build{
+		On: []catalogue.BuildsOn{{Arg: "RUNTIME_BASE", Module: "mesh-tools", Artifact: "runtime"}},
+	}}}
+	if !standsOnModule(fresh, "mesh-tools", nil) {
+		t.Fatal("a manifest's own base was not read")
+	}
+}
+
+// A merge names a repository the way the forge does; a source is recorded the way a build was
+// asked for. The two meet on owner/repo and branch, whichever form the record took.
+func TestAMergeMatchesTheSourcesBuiltFromIt(t *testing.T) {
+	m := link.SourceMoved{Owner: "novox", Repo: "mesh-controller", Base: "main",
+		CloneURL: "http://forge.internal:20000/novox/mesh-controller.git"}
+	for _, s := range []inventory.Source{
+		{Repository: "http://forge.internal:20000/novox/mesh-controller.git", Ref: "main"},
+		{Repository: "novox/mesh-controller", Seat: "git", Ref: ""},
+		{Repository: "https://elsewhere.example/novox/mesh-controller", Ref: "main"},
+	} {
+		if !sourceIs(s, m) {
+			t.Errorf("%+v was not matched by the merge", s)
+		}
+	}
+	for _, s := range []inventory.Source{
+		{Repository: "novox/mesh-host", Seat: "git"},
+		{Repository: "http://forge.internal:20000/novox/mesh-controller.git", Ref: "release"},
+	} {
+		if sourceIs(s, m) {
+			t.Errorf("%+v was matched by a merge that is not its", s)
+		}
+	}
+}
+
+// A merge made before the source was last seen is history: it does not move the source, and a
+// merge that says nothing about when it was made is taken as news.
+func TestAMergeOlderThanTheLastLookIsHistory(t *testing.T) {
+	seen := time.Date(2026, 9, 28, 3, 0, 0, 0, time.UTC)
+	if !isHistory("2026-09-28T02:00:00Z", seen) {
+		t.Fatal("an older merge was taken as news")
+	}
+	if isHistory("2026-09-28T04:00:00Z", seen) {
+		t.Fatal("a newer merge was taken as history")
+	}
+	if isHistory("", seen) || isHistory("2026-09-28T02:00:00Z", time.Time{}) {
+		t.Fatal("a merge or a source with no time on it was refused")
+	}
+}
+
+// A module as the catalogue holds it: built from a repository, at a directory inside it.
+func fromRepo(module, repository, path string) inventory.Entry {
+	return inventory.Entry{
+		Manifest: catalogue.Manifest{Module: module},
+		Source:   inventory.Source{Repository: repository, Path: path, Ref: "main"},
+	}
+}
+
+// A merge rebuilds the modules whose own directories it changed, and everything when what it changed
+// is shared. One repository holding many modules is the ordinary case here, and rebuilding all of
+// them for a change to one is what exhausted a registry's pull limit the first night this ran.
+func TestAMergeRebuildsTheModulesItChanged(t *testing.T) {
+	const repo = "http://forge.internal:20000/novox/mesh-catalog.git"
+	gitea := fromRepo("gitea", repo, "modules/gitea")
+	keycloak := fromRepo("keycloak", repo, "modules/keycloak")
+	known := []inventory.Entry{gitea, keycloak, fromRepo("plex", repo, "modules/plex")}
+	candidates := []inventory.Entry{gitea, keycloak}
+	merge := func(paths []string, truncated bool) link.SourceMoved {
+		return link.SourceMoved{Owner: "novox", Repo: "mesh-catalog", Base: "main",
+			Paths: paths, PathsTruncated: truncated}
+	}
+	named := func(entries []inventory.Entry) string {
+		var names []string
+		for _, e := range entries {
+			names = append(names, e.Manifest.Module)
+		}
+		return strings.Join(names, ",")
+	}
+	for _, c := range []struct {
+		what string
+		m    link.SourceMoved
+		want string
+	}{
+		{"one module's own files", merge([]string{"modules/gitea/index.ts", "modules/gitea/client.ts"}, false), "gitea"},
+		{"two modules' files", merge([]string{"modules/gitea/index.ts", "modules/keycloak/module.json"}, false), "gitea,keycloak"},
+		{"a file they share", merge([]string{"tsconfig.json"}, false), "gitea,keycloak"},
+		{"a module the mesh does not hold", merge([]string{"modules/plex/index.ts"}, false), ""},
+		{"nothing said about the files", merge(nil, false), "gitea,keycloak"},
+		{"more files than were listed", merge([]string{"modules/gitea/index.ts"}, true), "gitea,keycloak"},
+		// novox/hq issue 252: a module the mesh has never registered is still a module, when the merge
+		// shows it is one — and a directory that may be shared code is still shared.
+		{"a new module beside a held one", merge([]string{"modules/gitea/x", "modules/newmod/module.json"}, false), "gitea"},
+		{"a new module's other files", merge([]string{"modules/newmod/index.ts", "modules/newmod/module.json"}, false), ""},
+		{"a module removed", merge([]string{"modules/gone/module.json"}, false), ""},
+		{"a directory with no manifest", merge([]string{"modules/lib/x.go"}, false), "gitea,keycloak"},
+		{"a file directly among the modules", merge([]string{"modules/README.md"}, false), "gitea,keycloak"},
+		{"the root's files still", merge([]string{"tsconfig.json"}, false), "gitea,keycloak"},
+	} {
+		if got := named(whatTheMergeTouched(candidates, known, c.m)); got != c.want {
+			t.Errorf("%s: rebuilt %q, wanted %q", c.what, got, c.want)
+		}
+	}
+}
+
+// A module whose recipe packages source from another repository is affected when that repository
+// moves — the manifest the mesh keeps says nothing about it, so the record of what the build read is
+// the only thing that can say so.
+func TestAModuleIsAffectedByTheRepositoryItPackages(t *testing.T) {
+	m := link.SourceMoved{Owner: "novox", Repo: "mesh-controller", Base: "main",
+		CloneURL: "http://forge.internal:20000/novox/mesh-controller.git"}
+	for _, read := range [][]inventory.ReadRepository{
+		{{Repository: "http://forge.internal:20000/novox/mesh-controller.git", Ref: "main"}},
+		{{Repository: "novox/mesh-controller"}},
+		{{Repository: "https://elsewhere.example/novox/other"}, {Repository: "novox/mesh-controller.git", Ref: "main"}},
+	} {
+		if !readsFrom(read, m) {
+			t.Errorf("%+v was not matched by the merge", read)
+		}
+	}
+	for _, read := range [][]inventory.ReadRepository{
+		nil,
+		{{Repository: "novox/mesh-host", Ref: "main"}},
+		{{Repository: "novox/mesh-controller", Ref: "release"}},
+	} {
+		if readsFrom(read, m) {
+			t.Errorf("%+v was matched by a merge that is not its", read)
+		}
+	}
+}
+
+// What a module handed over by hand records about where it came from, and what is refused.
+func TestWhatAHandedOverModuleRecordsAboutItsSource(t *testing.T) {
+	// The whole location: a repository on the mesh's own forge, the directory inside it, the branch
+	// and the commit the manifest was read at.
+	from, err := whereItComesFrom("novox/mesh-catalog", "main", "c0ffee", "modules/gitea", true)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if from.Path != "modules/gitea" || from.Seat != "git" || from.Repository != "novox/mesh-catalog" {
+		t.Fatalf("the source records as %+v", from)
+	}
+	// A manifest with no provenance at all is legitimate: fixing something in a hurry.
+	if from, err := whereItComesFrom("", "", "", "", false); err != nil || !reflect.DeepEqual(from, inventory.Source{}) {
+		t.Fatalf("a manifest handed over with no provenance was refused: %+v, %v", from, err)
+	}
+	for _, c := range []struct {
+		what                          string
+		repository, ref, commit, path string
+		self                          bool
+	}{
+		{what: "a source with no commit", repository: "novox/mesh-catalog", commit: ""},
+		{what: "a commit with no source", commit: "c0ffee"},
+		{what: "a directory inside nothing", path: "modules/gitea"},
+		{what: "a forge holding nothing", self: true},
+		{what: "an address given as a path on the forge", repository: "http://forge.internal:20000/novox/x.git", commit: "c0ffee", self: true},
+	} {
+		if _, err := whereItComesFrom(c.repository, c.ref, c.commit, c.path, c.self); err == nil {
+			t.Errorf("%s was recorded as a source", c.what)
+		}
+	}
+}
+
+// novox/hq 04-ISSUES/215: a module once built at a commit still follows its branch — a merge into it
+// matches the module, and a plan re-asks the branch, not the old commit.
+func TestAModuleBuiltAtACommitStillFollowsItsBranch(t *testing.T) {
+	m := link.SourceMoved{Owner: "novox", Repo: "mesh-catalog", Base: "main"}
+	pinned := inventory.Source{Repository: "novox/mesh-catalog", Seat: "git", Ref: "9c97a8a"}
+	if !sourceIs(pinned, m) {
+		t.Error("a module whose record names a commit is left out of a merge into its branch")
+	}
+	full := inventory.Source{Repository: "novox/mesh-catalog", Seat: "git", Ref: "9c97a8a1d2c3b4a5f60718293a4b5c6d7e8f9012"}
+	if !sourceIs(full, m) {
+		t.Error("a full commit hash is read as a branch")
+	}
+	if got := followedBranch("9c97a8a"); got != "" {
+		t.Errorf("a plan would re-ask the old commit %q", got)
+	}
+	if got := followedBranch("release"); got != "release" {
+		t.Errorf("a branch is not followed as named: %q", got)
+	}
+	// A module that follows another branch is still not this merge's.
+	if sourceIs(inventory.Source{Repository: "novox/mesh-catalog", Seat: "git", Ref: "release"}, m) {
+		t.Error("a module following another branch was matched")
+	}
+}

cmd/mesh-controller/plan_retry.go

@@ -0,0 +1,401 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A plan follows what is done to the build queue (novox/hq ADR 0219).
+//
+// Two things a person does to builds by hand would otherwise leave a plan saying something untrue:
+//
+//   - **Pausing the build seat.** A plan whose builds wait in the queue of a seat whose every holder
+//     is paused is not late — nothing will take its asks until somebody resumes — and saying LATE
+//     sends a reader looking for a fault that is a decision. It says what it waits on instead.
+//   - **A build that failed, been cancelled or killed, and is asked again.** `plans retry` re-asks a
+//     failed plan's failed modules and the plan goes on from that tier as if they had built the
+//     first time; `rebuild` of a module a plan holds unbuilt joins that plan rather than running
+//     beside it — beside it, the plan would either ask it again or stay failed on an outcome the
+//     rebuild has already replaced.
+
+// pauseView is whether the build seat takes work: the holders that said they are paused, and
+// whether that is every holder.
+type pauseView struct {
+	Nodes []string
+	All   bool
+}
+
+// pausedWaiting is what a plan waits on when the build seat is paused under it, or false: a plan
+// building, whose current tier has an ask outstanding, while every holder of the seat is paused.
+func pausedWaiting(p inventory.Plan, pause pauseView, now time.Time) (string, bool) {
+	if p.State != inventory.PlanBuilding || !pause.All || len(pause.Nodes) == 0 || p.Tier >= len(p.Tiers) {
+		return "", false
+	}
+	var asked *time.Time
+	for _, m := range p.Tiers[p.Tier] {
+		if s := p.Modules[m]; s != nil && s.State == "asked" && s.AskedAt != nil {
+			if asked == nil || s.AskedAt.Before(*asked) {
+				asked = s.AskedAt
+			}
+		}
+	}
+	if asked == nil {
+		return "", false
+	}
+	waited := now.Sub(*asked)
+	said := fmt.Sprintf("waiting: the build seat is paused on %s (asked %s ago)",
+		strings.Join(pause.Nodes, ", "), waited.Round(time.Second))
+	// Not late — a pause is a decision — but a pause forgotten is a plan that never moves, so one held
+	// longer than a day is named.
+	if waited > pausedTooLong {
+		said += " — PAUSED OVER A DAY: `resume` takes builds again"
+	}
+	return said, true
+}
+
+// pausedTooLong is how long a plan may wait on a paused build seat before it says the pause is long.
+const pausedTooLong = 24 * time.Hour
+
+// awaitsABuild is whether any open plan has an ask outstanding in its current tier — the only case
+// where the seat being paused changes what a plan says.
+func awaitsABuild(plans []inventory.Plan) bool {
+	for _, p := range plans {
+		if p.State != inventory.PlanBuilding || p.Tier >= len(p.Tiers) {
+			continue
+		}
+		for _, m := range p.Tiers[p.Tier] {
+			if s := p.Modules[m]; s != nil && s.State == "asked" {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// buildSeatPause reads whether the build seat's holders take work, from what each last said on the
+// bus (link.HolderState) — read only when a plan waits on a build, so a mesh with nothing building
+// does not dial the bus to say so. Anything unreadable is said and read as not paused: a plan then
+// reads as late, which is what it said before this existed.
+func buildSeatPause(ctx context.Context, inv *inventory.Inventory, plans []inventory.Plan) pauseView {
+	if !awaitsABuild(plans) {
+		return pauseView{}
+	}
+	entries, err := inv.Catalogued(ctx)
+	if err != nil {
+		return pauseView{}
+	}
+	seat := buildSeatAmong(entries)
+	holders := holdersAmong(entries, seat)
+	if len(holders) == 0 {
+		return pauseView{}
+	}
+	address, err := broker.BusAddress()
+	if err != nil {
+		return pauseView{}
+	}
+	js, err := broker.Dial(address)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not reach the bus to read whether the build seat is paused: %v\n", err)
+		return pauseView{}
+	}
+	defer js.Close()
+	said, err := link.PausedSaid(js, seat, holders)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not read whether the build seat is paused: %v\n", err)
+		return pauseView{}
+	}
+	return pauseOf(holders, said)
+}
+
+// pauseOf is the view from what each holder said.
+func pauseOf(holders []string, said map[string]link.HolderState) pauseView {
+	var v pauseView
+	for _, n := range holders {
+		if said[n].Paused {
+			v.Nodes = append(v.Nodes, n)
+		}
+	}
+	sort.Strings(v.Nodes)
+	v.All = len(holders) > 0 && len(v.Nodes) == len(holders)
+	return v
+}
+
+// failedIn is the modules of a plan's current tier that failed to build, sorted.
+func failedIn(p inventory.Plan) []string {
+	if p.Tier >= len(p.Tiers) {
+		return nil
+	}
+	var out []string
+	for _, m := range p.Tiers[p.Tier] {
+		if s := p.Modules[m]; s != nil && s.State == "failed" {
+			out = append(out, m)
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+// newerOpenPlan is an open plan of the same repository and branch made after this one: the plan
+// that holds what this one held now (issue 254, ADR 0218).
+func newerOpenPlan(p inventory.Plan, plans []inventory.Plan) (inventory.Plan, bool) {
+	for _, q := range plans {
+		if q.ID == p.ID || !q.Open() || !strings.EqualFold(q.Repository, p.Repository) ||
+			(p.Branch != "" && q.Branch != "" && p.Branch != q.Branch) || !q.Created.After(p.Created) {
+			continue
+		}
+		return q, true
+	}
+	return inventory.Plan{}, false
+}
+
+// retryRefusal is why a plan cannot be retried, or nothing.
+func retryRefusal(p inventory.Plan, plans []inventory.Plan) error {
+	switch {
+	case p.State == inventory.PlanDone:
+		return fmt.Errorf("%s is done; there is nothing to retry", p.ID)
+	case p.State == inventory.PlanSuperseded:
+		return fmt.Errorf("%s was %s — what it had not built is in that plan", p.ID, p.Note)
+	case p.Open():
+		return fmt.Errorf("%s is still %s; nothing in it failed to retry — `rebuild <module>` asks one module again", p.ID, p.State)
+	}
+	if len(failedIn(p)) > 0 {
+		if q, found := newerOpenPlan(p, plans); found {
+			return fmt.Errorf("%s supersedes it: a newer merge of %s (%s at %s) is open, and retrying %s would build "+
+				"what that one replaced", q.ID, q.Repository, q.ID, short(q.Commit), p.ID)
+		}
+		return nil
+	}
+	stopped := stoppedRollouts(p)
+	if len(stopped) == 0 {
+		return fmt.Errorf("nothing in tier %d of %s failed to build or stopped rolling out — it stopped at: %s",
+			p.Tier, p.ID, p.Note)
+	}
+	// **A rollout is retried unless the module has moved on**: a newer plan holding it sends — or
+	// sent — a newer build, and sending this one again would put the older build back on its machines.
+	for _, m := range stopped {
+		if q, found := newerPlanFor(m, p, plans); found {
+			return fmt.Errorf("%s has a newer plan, %s (%s, %s at %s): sending %s's build of it again would put "+
+				"the older build back", m, q.ID, q.State, q.Repository, short(q.Commit), p.ID)
+		}
+	}
+	return nil
+}
+
+// stoppedRollouts is the modules of a plan's current tier whose rollout stopped at its first machine
+// (issue 249, ADR 0218): built, sent to the first machine, never to the rest, and why it stopped kept.
+func stoppedRollouts(p inventory.Plan) []string {
+	if p.Tier >= len(p.Tiers) {
+		return nil
+	}
+	var out []string
+	for _, m := range p.Tiers[p.Tier] {
+		if s := p.Modules[m]; s != nil && s.State == "built" && s.FirstAt != nil && s.SentAt == nil &&
+			len(s.First) > 0 && s.Why != "" {
+			out = append(out, m)
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+// newerPlanFor is a plan made after this one that holds the module, superseded ones aside.
+func newerPlanFor(module string, p inventory.Plan, plans []inventory.Plan) (inventory.Plan, bool) {
+	for _, q := range plans {
+		if q.ID == p.ID || q.State == inventory.PlanSuperseded || !q.Created.After(p.Created) {
+			continue
+		}
+		for _, tier := range q.Tiers {
+			for _, m := range tier {
+				if m == module {
+					return q, true
+				}
+			}
+		}
+	}
+	return inventory.Plan{}, false
+}
+
+// sendRollout sends machines what the mesh would send them now, answering the ones it sent. A
+// variable so a test of a retried rollout needs no machine.
+var sendRollout = sendToEach
+
+// resumed sets a failed plan building again once nothing in its tier is failed.
+func resumed(p *inventory.Plan, why string) {
+	if p.State == inventory.PlanFailed && len(failedIn(*p)) == 0 {
+		p.State = inventory.PlanBuilding
+		p.Note = why
+	}
+}
+
+// retryPlan asks a failed plan's failed modules again, under new ids, and sets it building again at
+// that tier: what follows is the plan going on as if they had built the first time.
+func retryPlan(ctx context.Context, open *stores, id string) (string, error) {
+	inv := open.inventory
+	release, err := inv.HoldPlans(ctx, true)
+	if err != nil {
+		return "", err
+	}
+	defer release()
+	p, err := inv.PlanByID(ctx, id)
+	if err != nil {
+		return "", err
+	}
+	plans, err := inv.OpenPlans(ctx)
+	if err != nil {
+		return "", err
+	}
+	recent, err := inv.RecentPlans(ctx, 50)
+	if err != nil {
+		return "", err
+	}
+	plans = append(plans, recent...)
+	if err := retryRefusal(p, plans); err != nil {
+		return "", err
+	}
+	if len(failedIn(p)) == 0 {
+		return retryRollouts(ctx, open, &p)
+	}
+	entries, err := inv.Catalogued(ctx)
+	if err != nil {
+		return "", err
+	}
+	byName := map[string]inventory.Entry{}
+	for _, e := range entries {
+		byName[e.Manifest.Module] = e
+	}
+	failed := failedIn(p)
+	var asked []string
+	for _, m := range failed {
+		askModule(ctx, &p, m, byName)
+		if s := p.Modules[m]; s.State == "asked" {
+			asked = append(asked, m+" as "+s.Build)
+		}
+	}
+	resumed(&p, fmt.Sprintf("tier %d retried by hand: %s asked again", p.Tier, strings.Join(failed, ", ")))
+	if err := inv.SavePlan(ctx, p); err != nil {
+		return "", err
+	}
+	if p.State != inventory.PlanBuilding {
+		return "", fmt.Errorf("%s could not be resumed: %s", p.ID, p.Note)
+	}
+	return fmt.Sprintf("%s retried at tier %d of %d: asked %s; the plan goes on from there as any plan does",
+		p.ID, p.Tier, len(p.Tiers), strings.Join(asked, ", ")), nil
+}
+
+// joinAPlan asks a module again for the plan that holds it unbuilt or failed, if one does: open plans
+// first, then the most recent failed one that nothing newer supersedes. Says whether it joined one.
+func joinAPlan(ctx context.Context, open *stores, module string) (bool, string, error) {
+	inv := open.inventory
+	release, err := inv.HoldPlans(ctx, true)
+	if err != nil {
+		return false, "", err
+	}
+	defer release()
+	openPlans, err := inv.OpenPlans(ctx)
+	if err != nil {
+		return false, "", err
+	}
+	recent, err := inv.RecentPlans(ctx, 20)
+	if err != nil {
+		return false, "", err
+	}
+	p, found := planHolding(module, openPlans, recent)
+	if !found {
+		return false, "", nil
+	}
+	entries, err := inv.Catalogued(ctx)
+	if err != nil {
+		return false, "", err
+	}
+	byName := map[string]inventory.Entry{}
+	for _, e := range entries {
+		byName[e.Manifest.Module] = e
+	}
+	was := p.State
+	askModule(ctx, &p, module, byName)
+	s := p.Modules[module]
+	resumed(&p, fmt.Sprintf("tier %d: %s rebuilt by hand", p.Tier, module))
+	if err := inv.SavePlan(ctx, p); err != nil {
+		return false, "", err
+	}
+	if s.State != "asked" {
+		return true, "", fmt.Errorf("%s could not be asked for %s: %s", module, p.ID, s.Why)
+	}
+	said := fmt.Sprintf("rebuild asked as %s, joining %s at tier %d (%s at %s): the plan takes this build as %s's outcome",
+		s.Build, p.ID, p.Tier, p.Repository, short(p.Commit), module)
+	switch {
+	case was == inventory.PlanFailed && p.State == inventory.PlanBuilding:
+		said += "; the plan had failed and builds again from this tier"
+	case was == inventory.PlanFailed:
+		said += "; the plan stays failed while " + strings.Join(failedIn(p), ", ") + " failed too — `plans retry " + p.ID + "` asks them"
+	}
+	return true, said, nil
+}
+
+// planHolding is the plan a rebuild of a module joins: an open plan whose current tier holds it not
+// yet built, else the newest failed plan whose current tier does, and that no open plan of its
+// repository supersedes.
+func planHolding(module string, openPlans, recent []inventory.Plan) (inventory.Plan, bool) {
+	holds := func(p inventory.Plan) bool {
+		if p.Tier >= len(p.Tiers) {
+			return false
+		}
+		for _, m := range p.Tiers[p.Tier] {
+			if m == module {
+				s := p.Modules[m]
+				return s == nil || s.State != "built"
+			}
+		}
+		return false
+	}
+	for _, p := range openPlans {
+		if holds(p) {
+			return p, true
+		}
+	}
+	sorted := append([]inventory.Plan(nil), recent...)
+	sort.SliceStable(sorted, func(i, j int) bool { return sorted[i].Created.After(sorted[j].Created) })
+	for _, p := range sorted {
+		if p.State != inventory.PlanFailed || !holds(p) {
+			continue
+		}
+		if _, superseded := newerOpenPlan(p, openPlans); superseded {
+			continue
+		}
+		return p, true
+	}
+	return inventory.Plan{}, false
+}
+
+// retryRollouts sends each module whose rollout stopped to the machines it was first sent to, again,
+// records that send as the first anew, and sets the plan rolling: from there it goes on as the plan
+// would have — the rest sent once those report they applied it, the next tier after (ADR 0218).
+func retryRollouts(ctx context.Context, open *stores, p *inventory.Plan) (string, error) {
+	var said []string
+	for _, m := range stoppedRollouts(*p) {
+		s := p.Modules[m]
+		sent, err := sendRollout(ctx, open, s.First)
+		if err != nil {
+			return "", fmt.Errorf("%s could not be sent to %s again, so %s stays failed: %w",
+				m, strings.Join(s.First, ", "), p.ID, err)
+		}
+		now := time.Now().UTC()
+		s.First, s.FirstAt, s.Why = sent, &now, ""
+		said = append(said, m+" to "+strings.Join(sent, ", "))
+	}
+	p.State = inventory.PlanRolling
+	p.Note = fmt.Sprintf("tier %d retried by hand; sent %s first again", p.Tier, strings.Join(said, "; "))
+	if err := open.inventory.SavePlan(ctx, *p); err != nil {
+		return "", err
+	}
+	return fmt.Sprintf("%s retried at tier %d of %d: sent %s first again; the rest follow once it reports it "+
+		"applied, as the plan would have", p.ID, p.Tier, len(p.Tiers), strings.Join(said, "; ")), nil
+}

cmd/mesh-controller/plan_test.go

@@ -0,0 +1,51 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// `plan` tells a person what a module would open and why, from the same `why` every listens
+// entry already carries for the firewall (novox/hq ADR 0007) — so deciding whether to assign a
+// module does not need reading its manifest first.
+func TestListensLinesShowWhatAModuleWouldOpenAndWhy(t *testing.T) {
+	m := catalogue.Manifest{Module: "minio", Listens: []catalogue.Listening{
+		{Port: 9000, From: catalogue.FromMesh, Why: "the S3 endpoint"},
+		{Port: 9001, From: catalogue.FromMesh},
+	}}
+	got := listensLines(m)
+	if len(got) != 2 {
+		t.Fatalf("two listens entries, got %d: %v", len(got), got)
+	}
+	if !strings.Contains(got[0], "9000/tcp") || !strings.Contains(got[0], "the S3 endpoint") {
+		t.Errorf("the port and its why did not both appear: %q", got[0])
+	}
+	if strings.Contains(got[1], "—") {
+		t.Errorf("a listens entry with no why should not print a dash: %q", got[1])
+	}
+	if !strings.Contains(got[1], "9001/tcp") {
+		t.Errorf("the port still appears without a why: %q", got[1])
+	}
+}
+
+func TestListensLinesAreEmptyForAModuleWithNothingToListenOn(t *testing.T) {
+	if got := listensLines(catalogue.Manifest{Module: "board"}); len(got) != 0 {
+		t.Errorf("a module with no listens should print nothing, got %v", got)
+	}
+}
+
+// `plan --show` says when a file is written into rather than over (novox/hq issue 128), or the
+// mesh's region of a hosts file reads as the whole file.
+func TestAFileWrittenIntoIsShownAsSuch(t *testing.T) {
+	region := shownAs(map[string]any{
+		"id": "mesh-wireguard.fact-node-names", "path": "/etc/hosts", "into": "block"})
+	if region != "mesh-wireguard.fact-node-names  /etc/hosts (written into, block)" {
+		t.Errorf("the region is shown as %q", region)
+	}
+	whole := shownAs(map[string]any{"id": "dnsmasq.fact-node-zones", "path": "/etc/mesh-resolver/nodes.conf"})
+	if strings.Contains(whole, "written into") {
+		t.Errorf("a whole file is shown as written into: %q", whole)
+	}
+}

cmd/mesh-controller/plans_one_at_a_time_test.go

@@ -0,0 +1,47 @@
+package main
+
+import (
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// novox/hq issue 213: for the moment a machine hands its controller over, the container and the
+// process both run the plan timer on one store. Only the one holding the plans moves them; the other
+// leaves them alone, and moves them once they are let go.
+func TestAControllerLeavesThePlansToTheOneHoldingThem(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	now := time.Now().UTC()
+	// Every tier done: the next step is the plan's last, and needs nothing but the store.
+	plan := inventory.Plan{ID: "plan-213", Repository: "r", Commit: "abc", Created: now, Updated: now,
+		State: inventory.PlanRolling, Tier: 1, Tiers: [][]string{{"app"}},
+		Modules: map[string]*inventory.PlanModule{"app": {State: "built"}}}
+	if err := open.inventory.SavePlan(ctx, plan); err != nil {
+		t.Fatal(err)
+	}
+
+	// The other controller: its own connections to the same store, holding the plans.
+	other, err := inventory.Open(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(other.Close)
+	release, err := other.HoldPlans(ctx, false)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(release) // before the close above: a pool waits for a connection still held
+
+	advancePlans(ctx, open)
+	if p, err := open.inventory.PlanByID(ctx, "plan-213"); err != nil || !p.Open() {
+		t.Fatalf("a controller moved a plan another held: %+v %v", p, err)
+	}
+
+	release()
+	advancePlans(ctx, open)
+	if p, err := open.inventory.PlanByID(ctx, "plan-213"); err != nil || p.State != inventory.PlanDone {
+		t.Fatalf("the plan did not move once it was let go: %+v %v", p, err)
+	}
+}

cmd/mesh-controller/push_test.go

@@ -0,0 +1,82 @@
+package main
+
+import (
+	"errors"
+	"strings"
+	"testing"
+)
+
+// One machine that cannot be worked out is not a reason to leave the mesh unconverged.
+//
+// A whole-mesh push refused outright the moment any single node failed to resolve, so a module on
+// the anchor requiring a provision nobody had assigned a provider for stopped every OTHER machine
+// from being sent anything — machines with no relation to the fault, and nothing wrong with them.
+// The failure and the punishment were on different machines.
+//
+// It is the same rule an un-hostable module already follows one level down (a92c11b: one module on
+// the wrong machine no longer refuses the whole node), applied one level up.
+func TestOneUnresolvableNodeStillLetsTheRestBeSent(t *testing.T) {
+	sending, refusals := composeEach(
+		[]string{"anchor", "home-server", "laptop"}, numbered(),
+		func(node string) (sendable, error) {
+			if node == "anchor" {
+				return sendable{}, errors.New(`nothing provides "acme-ca", wanted by route-proxy`)
+			}
+			return sendable{Resources: []map[string]any{{"id": node + ".thing"}}}, nil
+		})
+
+	var told []string
+	for _, s := range sending {
+		told = append(told, s.node)
+	}
+	if strings.Join(told, ",") != "home-server,laptop" {
+		t.Errorf("a machine with nothing wrong with it was not sent: %v", told)
+	}
+	if len(refusals) != 1 || !strings.Contains(refusals[0], "anchor") ||
+		!strings.Contains(refusals[0], "acme-ca") {
+		t.Errorf("the machine that could not be worked out was not named with its reason: %v",
+			refusals)
+	}
+}
+
+// A machine whose declaration composes to nothing is SENT the empty declaration, not skipped
+// (novox/hq issue 127): it may have held something before, and only sending the empty
+// declaration tells it to drop what the mesh owned. It is never a refusal.
+func TestAnEmptyDeclarationIsSentSoTheNodeDropsWhatItHeld(t *testing.T) {
+	sending, refusals := composeEach([]string{"spare"}, numbered(),
+		func(string) (sendable, error) { return sendable{}, nil })
+	if len(sending) != 1 || len(refusals) != 0 {
+		t.Errorf("an empty declaration must be sent, not skipped or refused: %v / %v", sending, refusals)
+	}
+}
+
+// A push that skipped a machine still ends badly, and says what was sent.
+//
+// **Skipping is not succeeding.** The mesh is not in the state somebody asked for, so the command
+// exits non-zero — but it says how many machines it did reach, because the old message ("nothing
+// was sent") was the very claim that had become untrue.
+func TestASkippedMachineIsStillAnError(t *testing.T) {
+	if err := couldNotBeResolved(nil, 3); err != nil {
+		t.Fatalf("a push that resolved every machine reported a problem: %v", err)
+	}
+
+	err := couldNotBeResolved([]string{"anchor:\nnothing provides \"acme-ca\""}, 2)
+	if err == nil {
+		t.Fatal("a push that could not work out a machine reported success")
+	}
+	said := err.Error()
+	if strings.Contains(said, "nothing was sent") {
+		t.Errorf("the push says nothing was sent, and it sent two machines: %q", said)
+	}
+	for _, want := range []string{"anchor", "acme-ca", "2 other node(s) were"} {
+		if !strings.Contains(said, want) {
+			t.Errorf("the refusal does not say %q: %q", want, said)
+		}
+	}
+}
+
+// numbered is an allotter for tests: one higher per call, as the inventory's is per machine.
+func numbered() func(string) (int64, error) {
+	var n int64
+	return func(string) (int64, error) { n++; return n, nil }
+}

cmd/mesh-controller/queue.go

@@ -0,0 +1,709 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"os"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/nats-io/nats.go"
+	"github.com/nats-io/nats.go/jetstream"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// The build queue, controlled by hand (novox/hq ADR 0219).
+//
+// Seen whole — what waits, what runs where and for how long, what was handed out as often as it
+// may be and never settled — and changed through the controller: an ask cancelled, the queue
+// cleared, a module rebuilt, a build replayed. What one machine is running is that machine's
+// holder's to end or pause, and the controller asks it (`kill`, `pause`, `resume`).
+//
+// **Everything that drops an ask leaves a failed outcome for it**, taken in exactly as a build that
+// failed is (takeIn, then the plan): a plan waiting on an ask a person removed fails, saying so,
+// rather than waiting for ever on an answer nobody will give.
+
+// holderAsks is how long a holder's verb is waited for; killAnswer how long its kill is — longer
+// than the holder's worst case (its two passes removing containers and its wait between, 50s).
+const (
+	holderAsks = 15 * time.Second
+	killAnswer = 75 * time.Second
+)
+
+// dialTheBus opens the controller's own connection, for a command that reads or changes the queue.
+func dialTheBus() (*broker.JetStream, error) {
+	address, err := broker.BusAddress()
+	if err != nil {
+		return nil, err
+	}
+	js, err := broker.Dial(address)
+	if err != nil {
+		return nil, fmt.Errorf("cannot reach the bus: %w", err)
+	}
+	return js, nil
+}
+
+// queueCommand prints every ask in the build seat's work queue.
+func queueCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("queue", flag.ContinueOnError)
+	asJSON := set.Bool("json", false, "the queue as JSON")
+	if _, err := parseAround(set, args); err != nil {
+		return err
+	}
+	js, err := dialTheBus()
+	if err != nil {
+		return err
+	}
+	defer js.Close()
+	seat := buildSeatHeld(ctx)
+	q, err := link.ReadQueue(ctx, js, seat)
+	if err != nil {
+		return err
+	}
+	if *asJSON {
+		body, err := json.MarshalIndent(q, "", "  ")
+		if err != nil {
+			return err
+		}
+		fmt.Println(string(body))
+		return nil
+	}
+	fmt.Print(queueText(q, time.Now()))
+	return nil
+}
+
+// queueText is the queue as a person reads it: a summary line, then each kind in queue order.
+// Never what an ask carries as `held` — that is every artifact the mesh has built.
+func queueText(q link.Queue, now time.Time) string {
+	var b strings.Builder
+	waiting, running, dead := q.Of(link.AskWaiting), q.Of(link.AskInFlight), q.Of(link.AskDead)
+	fmt.Fprintf(&b, "%s: %d waiting, %d in flight, %d dead\n", q.Seat, len(waiting), len(running), len(dead))
+	ago := func(t time.Time) string {
+		if t.IsZero() {
+			return "asked at an unknown time"
+		}
+		return "asked " + now.Sub(t).Round(time.Second).String() + " ago"
+	}
+	what := func(a link.QueuedAsk) string {
+		s := a.Repository
+		if a.Path != "" {
+			s += " at " + a.Path
+		}
+		if a.Ref != "" {
+			s += " on " + a.Ref
+		}
+		return s
+	}
+	if len(running) > 0 {
+		fmt.Fprintln(&b, "\nin flight:")
+		for _, a := range running {
+			where := "taken, not yet said where"
+			switch {
+			case a.On != "":
+				where = fmt.Sprintf("on %s for %s", a.On, now.Sub(a.Started).Round(time.Second))
+			case a.Was != "":
+				where = fmt.Sprintf("handed back by %s, to be handed out again", a.Was)
+			}
+			fmt.Fprintf(&b, "  %-26s %s — %s, %s (seq %d)\n", a.ID, what(a), where, ago(a.AskedAt), a.Seq)
+		}
+	}
+	if len(waiting) > 0 {
+		fmt.Fprintln(&b, "\nwaiting:")
+		for _, a := range waiting {
+			fmt.Fprintf(&b, "  %-26s %s — %s (seq %d)\n", a.ID, what(a), ago(a.AskedAt), a.Seq)
+		}
+	}
+	if len(dead) > 0 {
+		fmt.Fprintf(&b, "\ndead — handed out as often as the worker allows (%d) and never settled, still held:\n", q.MaxDeliver)
+		for _, a := range dead {
+			fmt.Fprintf(&b, "  %-26s %s — %s (seq %d)\n", a.ID, what(a), ago(a.AskedAt), a.Seq)
+		}
+	}
+	switch {
+	case len(q.Asks) == 0:
+		fmt.Fprintln(&b, "nothing is asked of it")
+	default:
+		fmt.Fprintln(&b, "\n`cancel <id>` drops a waiting or dead ask, `clear` every waiting one, `kill <id>` ends one in flight")
+	}
+	return b.String()
+}
+
+// cancelCommand drops one waiting or dead ask.
+func cancelCommand(ctx context.Context, args []string) error {
+	if len(args) != 1 {
+		return errors.New("cancel <build id>")
+	}
+	js, err := dialTheBus()
+	if err != nil {
+		return err
+	}
+	defer js.Close()
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	seat := buildSeatHeld(ctx)
+	q, err := link.ReadQueue(ctx, js, seat)
+	if err != nil {
+		return err
+	}
+	ask, found := q.Find(args[0])
+	if !found {
+		return fmt.Errorf("%s is not in the %s queue: it was built, cancelled, or never asked — `builds` says "+
+			"what came of it", args[0], seat)
+	}
+	said, err := cancelAsk(ctx, js, open, seat, ask)
+	if err != nil {
+		return err
+	}
+	fmt.Println(said)
+	return nil
+}
+
+// cancelAsk drops one ask from the queue and records it failed, cancelled by hand.
+//
+// **In this order, and each step for a reason** (novox/hq ADR 0219):
+//  1. an ask a machine says it is building is refused: deleting its message ends nothing a machine
+//     is running — that is `kill`, on the machine running it;
+//  2. its id goes into the seat's cancelled set, so a holder that fetches it from now on ends it;
+//  3. for a waiting ask, the worker is read again: one handed out since the queue was read was
+//     taken in the moment of cancelling, and the cancel is withdrawn and refused — the holder either
+//     read the mark first and ends it as cancelled, or is building it;
+//  4. for an ask the worker counts handed out that no machine is building — taken and not yet said,
+//     handed back after a restart, or past its deliveries while nobody pulls — the holder's own look
+//     at the set is waited out, and a start heard since withdraws and refuses the cancel: a holder
+//     that took it before the mark is building it, and only `kill` ends that;
+//  5. the message is deleted by its sequence;
+//  6. the failed outcome is taken in as any failed build's is, and the plan that asked fails.
+func cancelAsk(ctx context.Context, js *broker.JetStream, open *stores, seat string, ask link.QueuedAsk) (string, error) {
+	if ask.State == link.AskInFlight && ask.On != "" {
+		return "", fmt.Errorf("%s is in flight on %s: cancelling drops an ask nobody is building. `kill %s` "+
+			"ends the build where it runs", ask.ID, ask.On, ask.ID)
+	}
+	if err := link.MarkCancelled(ctx, js, seat, ask.ID); err != nil {
+		return "", err
+	}
+	withdraw := func() {
+		if err := link.UnmarkCancelled(ctx, js, seat, ask.ID); err != nil {
+			fmt.Fprintf(os.Stderr, "could not withdraw the cancel of %s: %v — a holder taking it ends it as cancelled\n", ask.ID, err)
+		}
+	}
+	worker, _ := broker.HolderConsumerFor("", "", broker.DeclaredSeat{Name: seat, Accepts: []string{"build"}})
+	switch ask.State {
+	case link.AskWaiting:
+		if taken, err := takenSince(js, worker, ask.Seq); err != nil {
+			withdraw()
+			return "", err
+		} else if taken {
+			withdraw()
+			return "", fmt.Errorf("%s was taken by a holder as it was cancelled, so the cancel is withdrawn. If the "+
+				"holder read it first it ends the ask as %s and its outcome says so; otherwise it is building — "+
+				"`queue` says which, and `kill %s` ends it", ask.ID, link.CancelledByHand, ask.ID)
+		}
+	case link.AskInFlight:
+		if started, err := startedSince(ctx, js, seat, ask); err != nil {
+			withdraw()
+			return "", err
+		} else if started != "" {
+			withdraw()
+			return "", fmt.Errorf("%s has started on %s since it was read, so the cancel is withdrawn: `kill %s` "+
+				"ends it there", ask.ID, started, ask.ID)
+		}
+	}
+	if err := js.Context().DeleteMsg(worker.Stream, ask.Seq); err != nil && !errors.Is(err, nats.ErrMsgNotFound) &&
+		!errors.Is(err, jetstream.ErrMsgNotFound) && !strings.Contains(err.Error(), "no message found") {
+		return "", fmt.Errorf("%s is marked cancelled and could not be deleted from the queue (seq %d): %w — a "+
+			"holder taking it ends it as cancelled", ask.ID, ask.Seq, err)
+	}
+	recordCancelled(ctx, open, ask)
+	return fmt.Sprintf("cancelled %s (%s, %s): deleted from the %s queue and recorded failed, %s — a plan "+
+		"that asked for it fails with that", ask.ID, ask.Repository, ask.State, seat, link.CancelledByHand), nil
+}
+
+// holderLooks is how long a cancel waits for a holder that took the ask before the mark to say it
+// started: longer than a holder's look at the cancelled set (3s) and its start that follows.
+var holderLooks = 5 * time.Second
+
+// startedSince waits out a holder's look at the cancelled set and says the machine that started the
+// ask since it was read, or nothing. A start it ended as cancelled comes with its outcome and is not
+// a start of a build.
+func startedSince(ctx context.Context, js *broker.JetStream, seat string, ask link.QueuedAsk) (string, error) {
+	select {
+	case <-ctx.Done():
+		return "", ctx.Err()
+	case <-time.After(holderLooks):
+	}
+	since := time.Now().Add(-7 * 24 * time.Hour)
+	if !ask.AskedAt.IsZero() {
+		since = ask.AskedAt.Add(-time.Minute)
+	}
+	heard, err := link.ReadBuildEvents(ctx, js, seat, since)
+	if err != nil {
+		return "", err
+	}
+	s, ok := heard.Started[ask.ID]
+	if !ok || heard.Outcomes[ask.ID] {
+		return "", nil
+	}
+	at, _ := time.Parse(time.RFC3339Nano, s.At)
+	if ask.Started.IsZero() || at.After(ask.Started) {
+		return s.On, nil
+	}
+	return "", nil
+}
+
+// takenSince is whether the worker has handed out the ask at this sequence.
+func takenSince(js *broker.JetStream, worker broker.Consumer, seq uint64) (bool, error) {
+	info, err := js.Context().ConsumerInfo(worker.Stream, worker.Name)
+	if errors.Is(err, nats.ErrConsumerNotFound) {
+		return false, nil
+	}
+	if err != nil {
+		return false, fmt.Errorf("cannot read the worker of the queue again: %w", err)
+	}
+	return info.Delivered.Stream >= seq, nil
+}
+
+// recordCancelled takes the failed outcome in the way the daemon takes in any build's: recorded,
+// then the plan that asked for it — by the id it asked with, or by repository and path.
+func recordCancelled(ctx context.Context, open *stores, ask link.QueuedAsk) {
+	r := ask.Request
+	result := link.BuildResult{ID: ask.ID, Repository: ask.Repository, Path: ask.Path, Ref: ask.Ref,
+		Source: r.Source, DryRun: r.DryRun, Failed: link.CancelledByHand}
+	_ = builds{open.inventory, open}.Built(ctx, result)
+}
+
+// clearCommand cancels every waiting ask, and with --dead every dead one too.
+func clearCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("clear", flag.ContinueOnError)
+	dead := set.Bool("dead", false, "the dead asks too")
+	if _, err := parseAround(set, args); err != nil {
+		return err
+	}
+	js, err := dialTheBus()
+	if err != nil {
+		return err
+	}
+	defer js.Close()
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	seat := buildSeatHeld(ctx)
+	said, err := clearQueue(ctx, js, open, seat, *dead)
+	fmt.Print(said)
+	return err
+}
+
+// clearQueue cancels what clear names, each as `cancel` would, and never anything in flight.
+func clearQueue(ctx context.Context, js *broker.JetStream, open *stores, seat string, dead bool) (string, error) {
+	q, err := link.ReadQueue(ctx, js, seat)
+	if err != nil {
+		return "", err
+	}
+	var b strings.Builder
+	cancelled, refused := 0, 0
+	for _, a := range q.Asks {
+		if a.State == link.AskInFlight || (a.State == link.AskDead && !dead) {
+			continue
+		}
+		said, err := cancelAsk(ctx, js, open, seat, a)
+		if err != nil {
+			refused++
+			fmt.Fprintf(&b, "  %s: %v\n", a.ID, err)
+			continue
+		}
+		cancelled++
+		fmt.Fprintf(&b, "  %s\n", said)
+	}
+	what := "waiting"
+	if dead {
+		what = "waiting and dead"
+	}
+	fmt.Fprintf(&b, "%d %s ask(s) cancelled", cancelled, what)
+	if refused > 0 {
+		fmt.Fprintf(&b, ", %d could not be", refused)
+	}
+	fmt.Fprintln(&b)
+	if n := len(q.Of(link.AskInFlight)); n > 0 {
+		fmt.Fprintf(&b, "%d in flight, left running: `kill <id>` ends one where it runs\n", n)
+	}
+	if n := len(q.Of(link.AskDead)); n > 0 && !dead {
+		fmt.Fprintf(&b, "%d dead, left: `clear --dead` cancels them too\n", n)
+	}
+	if refused > 0 {
+		return b.String(), fmt.Errorf("%d ask(s) could not be cancelled", refused)
+	}
+	return b.String(), nil
+}
+
+// rebuildCommand asks the module's current source again — or, given a build's id, that build's
+// repository, path and ref — under a new id.
+func rebuildCommand(ctx context.Context, args []string) error {
+	if len(args) != 1 {
+		return errors.New("rebuild <module | build id>")
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+	entries, err := inv.Catalogued(ctx)
+	if err != nil {
+		return err
+	}
+	var source buildSource
+	var path, ref, module string
+	if b, found, err := inv.BuildByID(ctx, args[0]); err != nil {
+		return err
+	} else if found {
+		source, module = sourceOfBuild(b, entries)
+		path, ref = b.Path, b.Ref
+		// **A build at a commit is rebuilt at what its module follows now** (novox/hq ADR 0219, issue
+		// 219): asked now, a rebuild of an old commit would be the newest ask of the module and roll
+		// that commit out over everything since. Building that commit again is `replay`, which says
+		// what it would do and refuses to register it while anything newer is asked or registered.
+		if e, known := entryNamed(entries, module); known && ref != "" && followedBranch(ref) == "" {
+			ref = followedBranch(e.Source.Ref)
+			follows := ref
+			if follows == "" {
+				follows = "the repository's default branch"
+			}
+			fmt.Printf("%s was built at commit %s; a rebuild asks what %s follows now, %s — `replay %s` "+
+				"builds that commit\n", b.ID, short(b.Ref), module, follows, b.ID)
+		}
+	} else if e, known := entryNamed(entries, args[0]); known {
+		// As a plan asks it: the branch it follows, never a commit a build once named (issue 215).
+		source = buildSource{Repository: e.Source.Repository, Seat: e.Source.Seat}
+		path, ref, module = e.Source.Path, followedBranch(e.Source.Ref), e.Manifest.Module
+	} else {
+		return fmt.Errorf("%s is neither a module the catalogue holds nor a build the mesh recorded", args[0])
+	}
+
+	// **A module a plan holds unbuilt, or failed, joins that plan** rather than running beside it: the
+	// plan would ask it again, or stay failed on an outcome a rebuild has replaced.
+	if module != "" {
+		joined, said, err := joinAPlan(ctx, open, module)
+		if err != nil {
+			return err
+		}
+		if joined {
+			fmt.Println(said)
+			return nil
+		}
+	}
+	id, err := askABuild(ctx, source, path, ref)
+	if err != nil {
+		return err
+	}
+	fmt.Printf("rebuild asked as %s\n", id)
+	return nil
+}
+
+// entryNamed is the catalogue's entry for a module.
+func entryNamed(entries []inventory.Entry, name string) (inventory.Entry, bool) {
+	for _, e := range entries {
+		if e.Manifest.Module == name {
+			return e, true
+		}
+	}
+	return inventory.Entry{}, false
+}
+
+// sourceOfBuild is where a recorded build's repository is asked from: the catalogued module's own
+// source when the build is of one — on its seat, so the outcome registers it as the mesh records it
+// (ADR 0111) — else the repository as it was cloned.
+func sourceOfBuild(b inventory.Build, entries []inventory.Entry) (buildSource, string) {
+	for _, e := range entries {
+		if (b.Module != "" && e.Manifest.Module == b.Module) ||
+			(b.Module == "" && repositoryMatches(e.Source.Repository, b.Repository) && e.Source.Path == b.Path) {
+			return buildSource{Repository: e.Source.Repository, Seat: e.Source.Seat}, e.Manifest.Module
+		}
+	}
+	return buildSource{Repository: b.Repository}, b.Module
+}
+
+// replayCommand asks a recorded build's repository and path again at the commit it built.
+func replayCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("replay", flag.ContinueOnError)
+	register := set.Bool("register", false, "register what it builds, as any build is")
+	older := set.Bool("older", false, "with --register: even though a newer build of the module is registered")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 1 {
+		return errors.New("replay <build id> [--register [--older]]")
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+	b, found, err := inv.BuildByID(ctx, positionals[0])
+	if err != nil {
+		return err
+	}
+	if !found {
+		return fmt.Errorf("no build %s is recorded", positionals[0])
+	}
+	entries, err := inv.Catalogued(ctx)
+	if err != nil {
+		return err
+	}
+	source, module := sourceOfBuild(b, entries)
+	var history []inventory.Build
+	if module != "" {
+		if history, err = inv.Builds(ctx, module, 100); err != nil {
+			return err
+		}
+	}
+	// What is asked of the module and not yet answered, when the replay would be registered.
+	var outstanding []string
+	if *register {
+		js, err := dialTheBus()
+		if err != nil {
+			return err
+		}
+		q, err := link.ReadQueue(ctx, js, buildSeatHeld(ctx))
+		js.Close()
+		if err != nil {
+			return err
+		}
+		plans, err := inv.OpenPlans(ctx)
+		if err != nil {
+			return err
+		}
+		outstanding = outstandingFor(module, b.Repository, b.Path, q, plans)
+	}
+	if err := replayRefusal(b, module, history, *register, *older, outstanding); err != nil {
+		return err
+	}
+	id, err := buildOneAsked(ctx, source, b.Path, b.Commit, 0, !*register)
+	if err != nil {
+		return err
+	}
+	fmt.Println(replaySaid(b, module, id, *register))
+	return nil
+}
+
+// replayRefusal is why a replay is not asked, or nothing (novox/hq ADR 0219, issue 207).
+//
+// A replay re-asks a recorded build at the commit it built, under a new id — and an id is when it
+// was asked, which is what orders builds of one module (issue 219). So a registered replay of an
+// older commit is newer than everything since, and would roll that older commit out as the module's
+// current version: what issue 207 recorded happening by accident. It is therefore a dry run unless
+// --register says otherwise, and --register is refused when a newer build of a different commit is
+// registered, unless --older says that is the point.
+//
+// **And refused while anything newer is outstanding**, --older or not: an ask of the module in the
+// queue, or an open plan holding it unbuilt. Asked now, the replay would be newer than those asks,
+// and their builds — made from newer source — would be recorded and never registered (issue 219
+// orders by ask), the plan waiting on them sending the older commit instead.
+func replayRefusal(b inventory.Build, module string, history []inventory.Build, register, older bool,
+	outstanding []string) error {
+	if b.Commit == "" {
+		return fmt.Errorf("%s recorded no commit — it failed before it knew what it was building, so there is "+
+			"nothing to replay; `rebuild %s` asks its repository, path and ref again", b.ID, b.ID)
+	}
+	if older && !register {
+		return errors.New("--older only says what --register may do; a dry run registers nothing")
+	}
+	if register && len(outstanding) > 0 {
+		return fmt.Errorf("%s is asked and not yet answered — %s — and a registered replay asked now would "+
+			"replace what those build (novox/hq issue 219). Wait for them, or `replay %s` without --register to look",
+			orNone(module), strings.Join(outstanding, "; "), b.ID)
+	}
+	if !register || older {
+		return nil
+	}
+	for _, h := range history {
+		if h.ID == b.ID || !h.Worked() || h.Commit == b.Commit {
+			continue
+		}
+		if h.AskedOrAt().After(b.AskedOrAt()) {
+			return fmt.Errorf("a newer build of %s is registered — %s, from %s — and registering a replay of %s "+
+				"would roll that older commit out as %s's current version (novox/hq issue 207). `replay %s` "+
+				"without --register looks at it; --register --older registers it anyway",
+				module, h.ID, short(h.Commit), short(b.Commit), module, b.ID)
+		}
+	}
+	return nil
+}
+
+// replaySaid is what a replay prints once asked: what it builds, and what becomes of the outcome.
+func replaySaid(b inventory.Build, module, id string, register bool) string {
+	what := b.Repository
+	if module != "" {
+		what = module
+	}
+	said := fmt.Sprintf("replaying %s (%s) at %s as %s", b.ID, what, short(b.Commit), id)
+	if !register {
+		return said + "\n  a dry run: the outcome is looked at and not taken in — nothing is recorded or " +
+			"registered, and nothing is sent. `builds --log " + id + "` follows it; `--register` registers it"
+	}
+	return said + "\n  registered when it is built, as the module's current version — newer than every build " +
+		"asked before now — and rolled out as its policy says. `builds --log " + id + "` follows it"
+}
+
+// killCommand finds the machine running a build and asks its holder to end it.
+func killCommand(ctx context.Context, args []string) error {
+	if len(args) != 1 {
+		return errors.New("kill <build id>")
+	}
+	id := args[0]
+	js, err := dialTheBus()
+	if err != nil {
+		return err
+	}
+	defer js.Close()
+	seat := buildSeatHeld(ctx)
+	since := time.Now().Add(-7 * 24 * time.Hour)
+	if at, ok := link.BuildAskedAt(id); ok {
+		since = at.Add(-time.Minute)
+	}
+	heard, err := link.ReadBuildEvents(ctx, js, seat, since)
+	if err != nil {
+		return err
+	}
+	started, ok := heard.Started[id]
+	if !ok {
+		return fmt.Errorf("no machine has said it started %s: if it waits in the queue, `cancel %s` drops it", id, id)
+	}
+	if heard.Outcomes[id] {
+		return fmt.Errorf("%s has already ended on %s — `builds` says how", id, started.On)
+	}
+	answer, err := link.AskSeatTool(ctx, js.Conn(), seat, "kill", started.On, map[string]string{"id": id}, killAnswer)
+	if err != nil {
+		return err
+	}
+	return printHolderAnswer(started.On, answer)
+}
+
+// pauseCommand asks one machine's holder, or every holder's, to pause or resume.
+func pauseCommand(ctx context.Context, verb string, args []string) error {
+	if len(args) > 1 {
+		return fmt.Errorf("%s [node]", verb)
+	}
+	js, err := dialTheBus()
+	if err != nil {
+		return err
+	}
+	defer js.Close()
+	seat := buildSeatHeld(ctx)
+	nodes := args
+	if len(nodes) == 0 {
+		if nodes, err = buildSeatHolders(ctx, seat); err != nil {
+			return err
+		}
+		if len(nodes) == 0 {
+			return fmt.Errorf("nothing holds %s, so there is nothing to %s", seat, verb)
+		}
+	}
+	failed := 0
+	for _, node := range nodes {
+		answer, err := link.AskSeatTool(ctx, js.Conn(), seat, verb, node, map[string]string{}, holderAsks)
+		if err != nil {
+			fmt.Printf("%s: %v\n", node, err)
+			failed++
+			continue
+		}
+		if err := printHolderAnswer(node, answer); err != nil {
+			failed++
+		}
+	}
+	if failed > 0 {
+		return fmt.Errorf("%d of %d machine(s) did not %s", failed, len(nodes), verb)
+	}
+	return nil
+}
+
+// printHolderAnswer prints what a holder said, or its refusal as the command's failure.
+func printHolderAnswer(node string, answer link.Answer) error {
+	if answer.Error != "" {
+		fmt.Printf("%s: %s\n", node, answer.Error)
+		return errors.New(answer.Error)
+	}
+	var said struct {
+		Said string `json:"said"`
+	}
+	if json.Unmarshal(answer.Result, &said) == nil && said.Said != "" {
+		fmt.Printf("%s: %s\n", node, said.Said)
+		return nil
+	}
+	fmt.Printf("%s: %s\n", node, string(answer.Result))
+	return nil
+}
+
+// buildSeatHolders is every machine an assigned module holding the build seat runs on.
+func buildSeatHolders(ctx context.Context, seat string) ([]string, error) {
+	open, err := openStores(ctx)
+	if err != nil {
+		return nil, err
+	}
+	defer open.Close()
+	entries, err := open.inventory.Catalogued(ctx)
+	if err != nil {
+		return nil, err
+	}
+	return holdersAmong(entries, seat), nil
+}
+
+// holdersAmong is the machines of every catalogued module claiming the seat, sorted, once each.
+func holdersAmong(entries []inventory.Entry, seat string) []string {
+	seen := map[string]bool{}
+	var out []string
+	for _, e := range entries {
+		if !e.Manifest.ClaimsSeat(seat) {
+			continue
+		}
+		for _, n := range e.On {
+			if !seen[n] {
+				seen[n] = true
+				out = append(out, n)
+			}
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+// outstandingFor is everything asked of a module and not yet answered: its asks in the build queue,
+// by repository and path, and every open plan holding it not yet built.
+func outstandingFor(module, repository, path string, q link.Queue, plans []inventory.Plan) []string {
+	var out []string
+	for _, a := range q.Asks {
+		if repositoryMatches(a.Repository, repository) && a.Path == path {
+			out = append(out, fmt.Sprintf("%s %s in the queue", a.ID, a.State))
+		}
+	}
+	if module == "" {
+		return out
+	}
+	for _, p := range plans {
+		if !p.Open() {
+			continue
+		}
+		for _, tier := range p.Tiers {
+			for _, m := range tier {
+				if m == module {
+					if st := p.Modules[m]; st == nil || st.State != "built" {
+						out = append(out, fmt.Sprintf("%s holds it not yet built", p.ID))
+					}
+				}
+			}
+		}
+	}
+	return out
+}

cmd/mesh-controller/queue_test.go

@@ -0,0 +1,772 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// The build queue, controlled by hand (novox/hq ADR 0219), and the plans that follow it.
+//
+//	docker run -d --rm --name bq-nats -p 14294:4222 nats:2.10-alpine -js
+//	make postgres PG_PORT=55566 PG_CONTAINER=bq-pg
+//	MESH_TEST_NATS=nats://127.0.0.1:14294 \
+//	MESH_TEST_POSTGRES='postgres://postgres:check@127.0.0.1:55566/postgres?sslmode=disable' \
+//	  go test ./cmd/mesh-controller/ -run 'Queue|Cancel|Clear|Retry|Rebuild|Replay|Paused'
+
+// asksRecorded makes every ask a plan makes return the next id in a row, and says which were asked.
+func asksRecorded(t *testing.T) *[]string {
+	t.Helper()
+	var asked []string
+	was := askABuild
+	askABuild = func(_ context.Context, source buildSource, path, ref string) (string, error) {
+		id := link.NewBuildID(time.Now().Add(time.Duration(len(asked)) * time.Millisecond))
+		asked = append(asked, source.Repository+"#"+id)
+		return id, nil
+	}
+	t.Cleanup(func() { askABuild = was })
+	return &asked
+}
+
+// twoTiers registers two modules, b standing on a, each with a source a plan asks.
+func twoTiers(t *testing.T, open *stores) {
+	t.Helper()
+	for _, name := range []string{"a", "b"} {
+		if err := open.inventory.RegisterModule(t.Context(), catalogue.Manifest{Module: name, Version: "1"},
+			inventory.Source{Repository: "novox/" + name, Seat: "git", Ref: "main", BuiltFrom: "c0ffee", Head: "c0ffee"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+}
+
+// A failed plan is retried: its failed module asked again under a new id, the plan building at that
+// tier, and when that build comes in the plan goes on and asks its next tier.
+func TestAFailedPlanIsRetriedAndGoesOnThroughItsLaterTiers(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	asked := asksRecorded(t)
+	twoTiers(t, open)
+	before := time.Now().UTC().Add(-time.Hour)
+	failed := inventory.Plan{ID: "plan-retry", Repository: "novox/a", Branch: "main", Commit: "c0ffee",
+		Created: before, State: inventory.PlanFailed, Tier: 0, Tiers: [][]string{{"a"}, {"b"}},
+		Note: "a failed to build in tier 0",
+		Modules: map[string]*inventory.PlanModule{"a": {State: "failed", AskedAt: &before, Build: "build-1",
+			Why: link.KilledByHand}}}
+	if err := open.inventory.SavePlan(ctx, failed); err != nil {
+		t.Fatal(err)
+	}
+
+	said, err := retryPlan(ctx, open, failed.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	p, err := open.inventory.PlanByID(ctx, failed.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	a := p.Modules["a"]
+	if p.State != inventory.PlanBuilding || a.State != "asked" || a.Build == "" || a.Build == "build-1" || a.Why != "" {
+		t.Fatalf("after retry the plan is %s and a is %+v", p.State, a)
+	}
+	if !strings.Contains(said, a.Build) {
+		t.Errorf("retry does not say the new id: %q", said)
+	}
+	if len(*asked) != 1 {
+		t.Fatalf("asked %v", *asked)
+	}
+
+	// The killed build's own late outcome is not this ask's; the new one's is, and tier 1 follows.
+	planBuilt(ctx, open, "a", "c0ffee", link.KilledByHand, before, "build-1")
+	if p, _ = open.inventory.PlanByID(ctx, failed.ID); p.State != inventory.PlanBuilding {
+		t.Fatalf("the old ask's outcome failed the retried plan: %s %q", p.State, p.Note)
+	}
+	asking, _ := link.BuildAskedAt(a.Build)
+	planBuilt(ctx, open, "a", "c0ffee", "", asking, a.Build)
+	p, err = open.inventory.PlanByID(ctx, failed.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if p.Tier != 1 || p.Modules["b"] == nil || p.Modules["b"].State != "asked" || p.Modules["b"].Build == "" {
+		t.Fatalf("the retried plan did not go on to tier 1: tier %d, %s, b %+v", p.Tier, p.State, p.Modules["b"])
+	}
+	if len(*asked) != 2 {
+		t.Fatalf("asked %v", *asked)
+	}
+}
+
+// What retry refuses, and says why.
+func TestRetryRefusesWhatItCannotResume(t *testing.T) {
+	at := time.Date(2026, 10, 5, 12, 0, 0, 0, time.UTC)
+	plan := func(id, state string, created time.Time) inventory.Plan {
+		return inventory.Plan{ID: id, Repository: "novox/mesh-catalog", Branch: "main", Commit: id + "c0ffee",
+			Created: created, State: state, Tiers: [][]string{{"a"}},
+			Modules: map[string]*inventory.PlanModule{"a": {State: "failed"}}}
+	}
+	failed := plan("plan-1", inventory.PlanFailed, at)
+	for _, c := range []struct {
+		p      inventory.Plan
+		others []inventory.Plan
+		says   string
+	}{
+		{plan("plan-d", inventory.PlanDone, at), nil, "is done"},
+		{plan("plan-s", inventory.PlanSuperseded, at), nil, "was"},
+		{plan("plan-o", inventory.PlanBuilding, at), nil, "still building"},
+		{failed, []inventory.Plan{plan("plan-2", inventory.PlanRolling, at.Add(time.Hour))}, "plan-2 supersedes it"},
+	} {
+		err := retryRefusal(c.p, c.others)
+		if err == nil || !strings.Contains(err.Error(), c.says) {
+			t.Errorf("%s: %v, wanted it to say %q", c.p.ID, err, c.says)
+		}
+	}
+	stopped := failed
+	stopped.Modules = map[string]*inventory.PlanModule{"a": {State: "built"}}
+	stopped.Note = "a stopped at its first machine"
+	if err := retryRefusal(stopped, nil); err == nil || !strings.Contains(err.Error(), "nothing in tier 0") {
+		t.Errorf("a plan with nothing failed to build was retried: %v", err)
+	}
+	// Another branch's newer plan, an older one, and a failed one do not supersede it.
+	other := plan("plan-3", inventory.PlanBuilding, at.Add(time.Hour))
+	other.Branch = "release"
+	if err := retryRefusal(failed, []inventory.Plan{other, plan("plan-0", inventory.PlanBuilding, at.Add(-time.Hour)),
+		plan("plan-4", inventory.PlanFailed, at.Add(time.Hour))}); err != nil {
+		t.Errorf("refused for a plan that does not supersede it: %v", err)
+	}
+}
+
+// `rebuild` of a module a failed plan holds joins that plan; one held by nothing runs alone.
+func TestARebuildJoinsThePlanHoldingTheModule(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	asked := asksRecorded(t)
+	twoTiers(t, open)
+	before := time.Now().UTC().Add(-time.Hour)
+	failed := inventory.Plan{ID: "plan-join", Repository: "novox/a", Branch: "main", Commit: "c0ffee",
+		Created: before, State: inventory.PlanFailed, Tiers: [][]string{{"a"}, {"b"}},
+		Note:    "a failed to build in tier 0",
+		Modules: map[string]*inventory.PlanModule{"a": {State: "failed", AskedAt: &before, Build: "build-1", Why: link.CancelledByHand}}}
+	if err := open.inventory.SavePlan(ctx, failed); err != nil {
+		t.Fatal(err)
+	}
+	if err := rebuildCommand(ctx, []string{"a"}); err != nil {
+		t.Fatal(err)
+	}
+	p, err := open.inventory.PlanByID(ctx, failed.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if p.State != inventory.PlanBuilding || p.Modules["a"].State != "asked" || p.Modules["a"].Build == "build-1" {
+		t.Fatalf("the rebuild did not join the failed plan: %s %+v", p.State, p.Modules["a"])
+	}
+	if len(*asked) != 1 {
+		t.Fatalf("asked %v", *asked)
+	}
+	// b is in a tier not yet reached: nothing holds it in its current tier, so it is asked alone.
+	if err := rebuildCommand(ctx, []string{"b"}); err != nil {
+		t.Fatal(err)
+	}
+	if p, _ = open.inventory.PlanByID(ctx, failed.ID); p.Modules["b"] != nil {
+		t.Fatalf("a module the plan has not reached joined it: %+v", p.Modules["b"])
+	}
+	if len(*asked) != 2 {
+		t.Fatalf("asked %v", *asked)
+	}
+}
+
+// A failed plan an open plan of its repository supersedes is not joined: the open one holds the module.
+func TestARebuildJoinsTheOpenPlanBeforeASupersededFailedOne(t *testing.T) {
+	at := time.Date(2026, 10, 5, 12, 0, 0, 0, time.UTC)
+	failed := inventory.Plan{ID: "plan-old", Repository: "novox/a", Branch: "main", Created: at, State: inventory.PlanFailed,
+		Tiers: [][]string{{"a"}}, Modules: map[string]*inventory.PlanModule{"a": {State: "failed"}}}
+	newer := inventory.Plan{ID: "plan-new", Repository: "novox/a", Branch: "main", Created: at.Add(time.Hour),
+		State: inventory.PlanBuilding, Tiers: [][]string{{"x"}, {"a"}}, Modules: map[string]*inventory.PlanModule{}}
+	if _, found := planHolding("a", []inventory.Plan{newer}, []inventory.Plan{newer, failed}); found {
+		t.Fatal("joined a failed plan a newer open one supersedes")
+	}
+	if p, found := planHolding("a", nil, []inventory.Plan{failed}); !found || p.ID != "plan-old" {
+		t.Fatalf("did not join the failed plan holding it: %v %s", found, p.ID)
+	}
+	built := failed
+	built.Modules = map[string]*inventory.PlanModule{"a": {State: "built"}}
+	if _, found := planHolding("a", nil, []inventory.Plan{built}); found {
+		t.Fatal("joined a plan that has the module built")
+	}
+}
+
+// replay is a dry run unless registered, and registering an older commit than one registered since
+// is refused unless --older says it is meant (novox/hq issue 207).
+func TestReplayRefusesToRollAnOlderCommitOutUnlessToldTo(t *testing.T) {
+	asked := time.Date(2026, 10, 5, 12, 0, 0, 0, time.UTC)
+	old := inventory.Build{ID: "build-old", Module: "a", Commit: "0ldc0mm1t", Asked: asked}
+	newer := inventory.Build{ID: "build-new", Module: "a", Commit: "n3wc0mm1t", Asked: asked.Add(time.Hour)}
+	history := []inventory.Build{newer, old}
+
+	if err := replayRefusal(old, "a", history, false, false, nil); err != nil {
+		t.Errorf("a dry run was refused: %v", err)
+	}
+	err := replayRefusal(old, "a", history, true, false, nil)
+	if err == nil || !strings.Contains(err.Error(), "build-new") || !strings.Contains(err.Error(), "--older") {
+		t.Errorf("registering an older commit than the one registered was not refused: %v", err)
+	}
+	if err := replayRefusal(old, "a", history, true, true, nil); err != nil {
+		t.Errorf("--register --older was refused: %v", err)
+	}
+	if err := replayRefusal(newer, "a", history, true, false, nil); err != nil {
+		t.Errorf("registering the newest build again was refused: %v", err)
+	}
+	// A newer build of the same commit, or one that failed, is not a newer version to roll back from.
+	same := inventory.Build{ID: "build-same", Module: "a", Commit: old.Commit, Asked: asked.Add(2 * time.Hour)}
+	broken := inventory.Build{ID: "build-broken", Module: "a", Failed: "no", Asked: asked.Add(3 * time.Hour)}
+	if err := replayRefusal(old, "a", []inventory.Build{broken, same, old}, true, false, nil); err != nil {
+		t.Errorf("refused for a newer build of the same commit or a failed one: %v", err)
+	}
+	if err := replayRefusal(inventory.Build{ID: "build-x", Failed: "clone"}, "", nil, false, false, nil); err == nil {
+		t.Error("a build that recorded no commit was replayed")
+	}
+	if err := replayRefusal(old, "a", history, false, true, nil); err == nil {
+		t.Error("--older without --register was taken")
+	}
+	// **Nothing newer outstanding**, --older or not: an ask of the module in the queue, or an open plan
+	// holding it unbuilt, would be replaced by a replay asked now (issue 219).
+	q := link.Queue{Asks: []link.QueuedAsk{
+		{ID: "build-queued", Repository: "https://forge.example/novox/a.git", State: link.AskWaiting},
+		{ID: "build-elsewhere", Repository: "https://forge.example/novox/b.git", State: link.AskWaiting},
+		{ID: "build-other-path", Repository: "https://forge.example/novox/a.git", Path: "sub", State: link.AskWaiting},
+	}}
+	plans := []inventory.Plan{
+		{ID: "plan-holds", State: inventory.PlanBuilding, Tiers: [][]string{{"x"}, {"a"}}, Modules: map[string]*inventory.PlanModule{}},
+		{ID: "plan-built", State: inventory.PlanRolling, Tiers: [][]string{{"a"}}, Modules: map[string]*inventory.PlanModule{"a": {State: "built"}}},
+		{ID: "plan-failed", State: inventory.PlanFailed, Tiers: [][]string{{"a"}}, Modules: map[string]*inventory.PlanModule{}},
+	}
+	outstanding := outstandingFor("a", "novox/a", "", q, plans)
+	if len(outstanding) != 2 || !strings.Contains(outstanding[0], "build-queued") || !strings.Contains(outstanding[1], "plan-holds") {
+		t.Fatalf("outstanding: %v", outstanding)
+	}
+	if err := replayRefusal(old, "a", history, true, true, outstanding); err == nil || !strings.Contains(err.Error(), "build-queued") {
+		t.Errorf("registered over an outstanding ask: %v", err)
+	}
+	if err := replayRefusal(old, "a", history, false, false, outstanding); err != nil {
+		t.Errorf("a dry run was refused for what is outstanding: %v", err)
+	}
+	if said := replaySaid(old, "a", "build-1", false); !strings.Contains(said, "dry run") || !strings.Contains(said, "--register") {
+		t.Errorf("a dry replay does not say what it is: %q", said)
+	}
+	if said := replaySaid(old, "a", "build-1", true); !strings.Contains(said, "registered") || !strings.Contains(said, "rolled out") {
+		t.Errorf("a registered replay does not say what it does: %q", said)
+	}
+}
+
+// A plan waiting on builds of a seat whose every holder is paused says so and is not late; a seat
+// paused on some holders only is not a reason the plan is waiting.
+func TestAPlanWaitingOnAPausedSeatSaysSoAndIsNotLate(t *testing.T) {
+	now := time.Date(2026, 10, 5, 12, 0, 0, 0, time.UTC)
+	asked := now.Add(-12 * time.Minute)
+	p := inventory.Plan{ID: "plan-p", Repository: "novox/a", Commit: "c0ffee", State: inventory.PlanBuilding,
+		Updated: now.Add(-2 * time.Hour), Tiers: [][]string{{"a"}},
+		Modules: map[string]*inventory.PlanModule{"a": {State: "asked", AskedAt: &asked}}}
+
+	all := pauseOf([]string{"g14", "ace"}, map[string]link.HolderState{"ace": {Paused: true}, "g14": {Paused: true}})
+	line := planLineWith(p, now, all)
+	if !strings.Contains(line, "waiting: the build seat is paused on ace, g14 (asked 12m0s ago)") || strings.Contains(line, "LATE") {
+		t.Errorf("a plan on a paused seat reads %q", line)
+	}
+	if st := planStatuses([]inventory.Plan{p}, now, all)[0]; st.Late || !strings.Contains(st.Waiting, "paused on ace, g14") {
+		t.Errorf("status --json says %+v", st)
+	}
+	if _, late := openPlans([]inventory.Plan{p}, all); late != 0 {
+		t.Errorf("a plan waiting on a paused seat counted late")
+	}
+
+	longAgo := now.Add(-25 * time.Hour)
+	forgotten := p
+	forgotten.Modules = map[string]*inventory.PlanModule{"a": {State: "asked", AskedAt: &longAgo}}
+	if line := planLineWith(forgotten, now, all); !strings.Contains(line, "PAUSED OVER A DAY") || strings.Contains(line, "LATE") {
+		t.Errorf("a plan paused over a day reads %q", line)
+	}
+
+	some := pauseOf([]string{"g14", "ace"}, map[string]link.HolderState{"ace": {Paused: true}})
+	if some.All || !reflect.DeepEqual(some.Nodes, []string{"ace"}) {
+		t.Fatalf("%+v", some)
+	}
+	if line := planLineWith(p, now, some); !strings.Contains(line, "LATE") {
+		t.Errorf("a seat paused on one holder of two made the plan not late: %q", line)
+	}
+	if st := planStatuses([]inventory.Plan{p}, now, some)[0]; !st.Late {
+		t.Errorf("status --json: %+v", st)
+	}
+	// A plan rolling out, or with nothing asked, is not waiting on the seat.
+	rolling := p
+	rolling.State = inventory.PlanRolling
+	if _, paused := pausedWaiting(rolling, all, now); paused {
+		t.Error("a rolling plan reads as waiting on the build seat")
+	}
+}
+
+// The queue's verbs are the controller seat's, each to the command it names.
+func TestTheQueueVerbsRunTheirCommands(t *testing.T) {
+	for _, c := range []struct {
+		verb string
+		args map[string]any
+		want []string
+	}{
+		{"queue", nil, []string{"queue"}},
+		{"cancel", map[string]any{"id": "build-1"}, []string{"cancel", "build-1"}},
+		{"clear", nil, []string{"clear"}},
+		{"clear", map[string]any{"dead": "true"}, []string{"clear", "--dead"}},
+		{"rebuild", map[string]any{"what": "gitea"}, []string{"rebuild", "gitea"}},
+		{"replay", map[string]any{"id": "build-1"}, []string{"replay", "build-1"}},
+		{"replay", map[string]any{"id": "build-1", "register": "true", "older": "true"}, []string{"replay", "build-1", "--register", "--older"}},
+		{"kill", map[string]any{"id": "build-1"}, []string{"kill", "build-1"}},
+		{"pause", nil, []string{"pause"}},
+		{"resume", map[string]any{"node": "ace"}, []string{"resume", "ace"}},
+		{"plans", map[string]any{"retry": "plan-1"}, []string{"plans", "retry", "plan-1"}},
+	} {
+		got, err := argvFor(c.verb, c.args)
+		if err != nil || !reflect.DeepEqual(got, c.want) {
+			t.Errorf("%s %v: %v %v, want %v", c.verb, c.args, got, err, c.want)
+		}
+	}
+	if _, err := argvFor("cancel", nil); err == nil {
+		t.Error("cancel without an id was taken")
+	}
+	declared := map[string]bool{}
+	for _, v := range catalogue.ControllerVerbs {
+		declared[v.Name] = true
+	}
+	for _, v := range []string{"queue", "cancel", "clear", "rebuild", "replay", "kill", "pause", "resume"} {
+		if !declared[v] {
+			t.Errorf("%s is not a verb of the controller seat", v)
+		}
+	}
+}
+
+// --- against a real bus -----------------------------------------------------------------------
+
+// aBuildQueue is the build seat's queue, worker and cancelled set on a real server, the controller
+// pointed at it, and a function that asks the seat one build.
+func aBuildQueue(t *testing.T) (*broker.JetStream, func(id, repository string) link.BuildRequest) {
+	t.Helper()
+	url := os.Getenv("MESH_TEST_NATS")
+	if url == "" {
+		t.Skip("MESH_TEST_NATS unset")
+	}
+	t.Setenv(broker.NATSVar, url)
+	js, err := broker.Dial(url)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(js.Close)
+	seat := broker.DeclaredSeat{Name: link.TheBuildMachine, Accepts: []string{"build"},
+		Emits: []string{"started", "built", "log.*", "paused.*"}}
+	if err := broker.AssertMeshStreams(js); err != nil {
+		t.Fatal(err)
+	}
+	_ = js.Context().DeleteStream("SEAT_NODE_BUILD_AGENT")
+	if err := broker.RaiseSeats(js, []broker.DeclaredSeat{seat}, map[string]broker.Holder{
+		link.TheBuildMachine: {Node: "anchor", Module: "build-agent"}}); err != nil {
+		t.Fatal(err)
+	}
+	if err := broker.RaiseCancelledSets(js, []broker.DeclaredSeat{seat}); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() {
+		_ = js.Context().DeleteStream("SEAT_NODE_BUILD_AGENT")
+		_ = js.Context().DeleteKeyValue(broker.CancelledSetName(link.TheBuildMachine))
+		_ = js.Context().PurgeStream(broker.EventsStream)
+	})
+	ask := func(id, repository string) link.BuildRequest {
+		r := link.BuildRequest{ID: id, Repository: repository, Held: map[string]string{"x/y": "secret-ish"}}
+		body, _ := json.Marshal(r)
+		if _, err := js.Context().Publish(link.BuildWork(), body); err != nil {
+			t.Fatal(err)
+		}
+		return r
+	}
+	return js, ask
+}
+
+// deadOne takes the oldest ask from the worker and hands it back as often as the worker allows.
+func deadOne(t *testing.T, js *broker.JetStream) {
+	t.Helper()
+	worker, _ := broker.HolderConsumerFor("", "", broker.DeclaredSeat{Name: link.TheBuildMachine, Accepts: []string{"build"}})
+	sub, err := js.Context().PullSubscribe(worker.Filters[0], worker.Name, nats.Bind(worker.Stream, worker.Name), nats.ManualAck())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = sub.Unsubscribe() }()
+	for i := 0; i < worker.MaxDeliver; i++ {
+		msgs, err := sub.Fetch(1, nats.MaxWait(3*time.Second))
+		if err != nil {
+			t.Fatalf("delivery %d: %v", i+1, err)
+		}
+		_ = msgs[0].Nak()
+	}
+	// One more pull, as a holder always has one waiting: the server finds the ask past its deliveries
+	// then, and stops counting it pending.
+	if msgs, _ := sub.Fetch(1, nats.MaxWait(time.Second)); len(msgs) > 0 {
+		t.Fatalf("an ask past its deliveries was delivered again")
+	}
+}
+
+// cancel drops a waiting ask from the bus and records it failed, cancelled by hand — and the plan
+// that asked for it, matched by the id, fails with it; an ask the plan's records name no module for
+// is still found.
+func TestCancelDeletesTheAskAndFailsThePlanThatAskedIt(t *testing.T) {
+	js, ask := aBuildQueue(t)
+	open := aMesh(t)
+	ctx := t.Context()
+	twoTiers(t, open)
+	id := link.NewBuildID(time.Now())
+	ask(id, "https://forge.example/novox/a.git")
+	asked, _ := link.BuildAskedAt(id)
+	plan := inventory.Plan{ID: "plan-cancel", Repository: "novox/a", Commit: "c0ffee", Created: asked,
+		State: inventory.PlanBuilding, Tiers: [][]string{{"a"}, {"b"}},
+		Modules: map[string]*inventory.PlanModule{"a": {State: "asked", AskedAt: &asked, Build: id}}}
+	if err := open.inventory.SavePlan(ctx, plan); err != nil {
+		t.Fatal(err)
+	}
+
+	q, err := link.ReadQueue(ctx, js, link.TheBuildMachine)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(q.Asks) != 1 || q.Asks[0].State != link.AskWaiting || q.Asks[0].ID != id {
+		t.Fatalf("the queue reads %+v", q)
+	}
+	if text := queueText(q, time.Now()); !strings.Contains(text, "1 waiting, 0 in flight, 0 dead") ||
+		strings.Contains(text, "secret-ish") {
+		t.Errorf("the queue says:\n%s", text)
+	}
+
+	if err := cancelCommand(ctx, []string{id}); err != nil {
+		t.Fatal(err)
+	}
+	if q, _ = link.ReadQueue(ctx, js, link.TheBuildMachine); len(q.Asks) != 0 {
+		t.Fatalf("the ask is still queued: %+v", q.Asks)
+	}
+	if cancelled, err := link.IsCancelled(js.Conn(), link.TheBuildMachine, id); err != nil || !cancelled {
+		t.Errorf("the cancelled set does not hold it: %v %v", cancelled, err)
+	}
+	b, found, err := open.inventory.BuildByID(ctx, id)
+	if err != nil || !found || b.Failed != link.CancelledByHand {
+		t.Fatalf("the cancel is recorded as %+v (%v %v)", b, found, err)
+	}
+	p, err := open.inventory.PlanByID(ctx, plan.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if p.State != inventory.PlanFailed || p.Modules["a"].State != "failed" || p.Modules["a"].Why != link.CancelledByHand {
+		t.Fatalf("the plan that asked is %s, a %+v", p.State, p.Modules["a"])
+	}
+	if err := cancelCommand(ctx, []string{id}); err == nil {
+		t.Error("an ask cancelled already was cancelled again")
+	}
+}
+
+// clear cancels every waiting ask and leaves the dead ones unless told, and never one in flight.
+func TestClearCancelsTheWaitingAndTheDeadOnlyWhenTold(t *testing.T) {
+	js, ask := aBuildQueue(t)
+	open := aMesh(t)
+	ctx := t.Context()
+	start := time.Now()
+	dead := link.NewBuildID(start)
+	ask(dead, "https://forge.example/novox/dead.git")
+	deadOne(t, js)
+	var waiting []string
+	for i := 1; i <= 2; i++ {
+		id := link.NewBuildID(start.Add(time.Duration(i) * time.Millisecond))
+		waiting = append(waiting, id)
+		ask(id, fmt.Sprintf("https://forge.example/novox/w%d.git", i))
+	}
+
+	q, err := link.ReadQueue(ctx, js, link.TheBuildMachine)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(q.Of(link.AskDead)) != 1 || len(q.Of(link.AskWaiting)) != 2 || q.MaxDeliver != 5 {
+		t.Fatalf("the queue reads %+v", q)
+	}
+	said, err := clearQueue(ctx, js, open, link.TheBuildMachine, false)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(said, "2 waiting ask(s) cancelled") || !strings.Contains(said, "1 dead, left") {
+		t.Errorf("clear said:\n%s", said)
+	}
+	q, _ = link.ReadQueue(ctx, js, link.TheBuildMachine)
+	if len(q.Asks) != 1 || q.Asks[0].ID != dead || q.Asks[0].State != link.AskDead {
+		t.Fatalf("after clear the queue is %+v", q.Asks)
+	}
+	if _, err := clearQueue(ctx, js, open, link.TheBuildMachine, true); err != nil {
+		t.Fatal(err)
+	}
+	if q, _ = link.ReadQueue(ctx, js, link.TheBuildMachine); len(q.Asks) != 0 {
+		t.Fatalf("clear --dead left %+v", q.Asks)
+	}
+	for _, id := range append(waiting, dead) {
+		if b, found, _ := open.inventory.BuildByID(ctx, id); !found || b.Failed != link.CancelledByHand {
+			t.Errorf("%s is recorded as %+v", id, b)
+		}
+	}
+}
+
+// An ask in flight is not cancelled: kill ends it where it runs.
+func TestCancelRefusesAnAskInFlight(t *testing.T) {
+	js, ask := aBuildQueue(t)
+	open := aMesh(t)
+	ctx := t.Context()
+	id := link.NewBuildID(time.Now())
+	ask(id, "https://forge.example/novox/a.git")
+	worker, _ := broker.HolderConsumerFor("", "", broker.DeclaredSeat{Name: link.TheBuildMachine, Accepts: []string{"build"}})
+	sub, err := js.Context().PullSubscribe(worker.Filters[0], worker.Name, nats.Bind(worker.Stream, worker.Name), nats.ManualAck())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = sub.Unsubscribe() }()
+	if _, err := sub.Fetch(1, nats.MaxWait(3*time.Second)); err != nil {
+		t.Fatal(err)
+	}
+	started, _ := json.Marshal(link.BuildStart{ID: id, On: "ace", At: time.Now().UTC().Format(time.RFC3339Nano)})
+	if _, err := js.Context().Publish(link.BuildStarted(), started); err != nil {
+		t.Fatal(err)
+	}
+	q, err := link.ReadQueue(ctx, js, link.TheBuildMachine)
+	if err != nil {
+		t.Fatal(err)
+	}
+	a, _ := q.Find(id)
+	if a.State != link.AskInFlight || a.On != "ace" {
+		t.Fatalf("the taken ask reads %+v", a)
+	}
+	_, err = cancelAsk(ctx, js, open, link.TheBuildMachine, a)
+	if err == nil || !strings.Contains(err.Error(), "kill "+id) {
+		t.Fatalf("an ask in flight was cancelled: %v", err)
+	}
+	if _, found, _ := open.inventory.BuildByID(ctx, id); found {
+		t.Error("a refused cancel recorded an outcome")
+	}
+}
+
+// kill finds the machine from the build's start and asks that machine's holder; pause asks the
+// machine named. Each prints what the holder answered, and a refusal is the command's failure.
+func TestKillAndPauseAskTheHolderOnTheMachine(t *testing.T) {
+	js, _ := aBuildQueue(t)
+	ctx := t.Context()
+	asked := map[string]string{}
+	handlers := map[string]link.ToolHandler{
+		"kill": func(_ context.Context, raw json.RawMessage) (any, error) {
+			var args struct{ ID string }
+			_ = json.Unmarshal(raw, &args)
+			asked["kill"] = args.ID
+			if args.ID != "build-running" {
+				return nil, fmt.Errorf("ace is not building %s", args.ID)
+			}
+			return map[string]any{"said": "killed " + args.ID}, nil
+		},
+		"pause": func(context.Context, json.RawMessage) (any, error) {
+			asked["pause"] = "ace"
+			return map[string]any{"said": "ace is paused"}, nil
+		},
+	}
+	stop, err := link.OverNATS{Conn: js.Conn()}.ServeNodeSeatTools(link.TheBuildMachine, "ace", handlers, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer stop()
+	for _, id := range []string{"build-running", "build-other"} {
+		started, _ := json.Marshal(link.BuildStart{ID: id, On: "ace", At: time.Now().UTC().Format(time.RFC3339Nano)})
+		if _, err := js.Context().Publish(link.BuildStarted(), started); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := killCommand(ctx, []string{"build-running"}); err != nil {
+		t.Fatal(err)
+	}
+	if asked["kill"] != "build-running" {
+		t.Fatalf("the holder was asked %v", asked)
+	}
+	if err := killCommand(ctx, []string{"build-other"}); err == nil {
+		t.Error("the holder's refusal was not the command's")
+	}
+	if err := killCommand(ctx, []string{"build-never"}); err == nil || !strings.Contains(err.Error(), "cancel build-never") {
+		t.Errorf("a build nobody started: %v", err)
+	}
+	if err := pauseCommand(ctx, "pause", []string{"ace"}); err != nil || asked["pause"] != "ace" {
+		t.Fatalf("pause: %v %v", err, asked)
+	}
+	if err := pauseCommand(ctx, "resume", []string{"g14"}); err == nil {
+		t.Error("a machine nothing answers on was resumed")
+	}
+}
+
+// A plan that stopped at its first machine is retried: the module sent to that machine again, the
+// send recorded as the first anew, and the plan goes on — unless a newer plan holds the module.
+func TestAPlanStoppedAtItsFirstMachineIsRetried(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	asked := asksRecorded(t)
+	twoTiers(t, open)
+	var sentTo [][]string
+	was := sendRollout
+	sendRollout = func(_ context.Context, _ *stores, names []string) ([]string, error) {
+		sentTo = append(sentTo, names)
+		return names, nil
+	}
+	t.Cleanup(func() { sendRollout = was })
+
+	long := time.Now().UTC().Add(-2 * time.Hour)
+	stopped := inventory.Plan{ID: "plan-rollout", Repository: "novox/a", Branch: "main", Commit: "c0ffee",
+		Created: long, State: inventory.PlanFailed, Tiers: [][]string{{"a"}, {"b"}},
+		Note: "a stopped at its first machine in tier 0: laptop refused what it was sent",
+		Modules: map[string]*inventory.PlanModule{"a": {State: "built", BuiltAt: &long, Commit: "c0ffee",
+			First: []string{"laptop"}, FirstAt: &long, Why: "laptop refused what it was sent"}}}
+	if err := open.inventory.SavePlan(ctx, stopped); err != nil {
+		t.Fatal(err)
+	}
+
+	// A newer plan holding a refuses it: sending the older build would put it back.
+	newer := inventory.Plan{ID: "plan-newer", Repository: "novox/other", Commit: "d00d", Created: long.Add(time.Hour),
+		State: inventory.PlanDone, Tiers: [][]string{{"a"}}, Modules: map[string]*inventory.PlanModule{}}
+	if err := open.inventory.SavePlan(ctx, newer); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := retryPlan(ctx, open, stopped.ID); err == nil || !strings.Contains(err.Error(), "plan-newer") {
+		t.Fatalf("retried under a newer plan: %v", err)
+	}
+	newer.State = inventory.PlanSuperseded
+	if err := open.inventory.SavePlan(ctx, newer); err != nil {
+		t.Fatal(err)
+	}
+
+	said, err := retryPlan(ctx, open, stopped.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(sentTo) != 1 || !reflect.DeepEqual(sentTo[0], []string{"laptop"}) || !strings.Contains(said, "laptop") {
+		t.Fatalf("sent %v; said %q", sentTo, said)
+	}
+	p, err := open.inventory.PlanByID(ctx, stopped.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	a := p.Modules["a"]
+	if p.State != inventory.PlanRolling || a.FirstAt == nil || !a.FirstAt.After(long) || a.Why != "" {
+		t.Fatalf("after retry the plan is %s, a %+v", p.State, a)
+	}
+	// And it goes on: a records (its policy sends nothing more), so the next tier is asked.
+	advancePlans(ctx, open)
+	if p, _ = open.inventory.PlanByID(ctx, stopped.ID); p.Tier != 1 || p.Modules["b"] == nil || p.Modules["b"].State != "asked" {
+		t.Fatalf("the retried plan did not go on: tier %d %s %+v", p.Tier, p.State, p.Modules["b"])
+	}
+	if len(*asked) != 1 {
+		t.Fatalf("asked %v", *asked)
+	}
+}
+
+// A plan module asked under an id is settled by that id's outcome alone: a replay or a rebuild beside
+// the plan, asked later, never answers it (novox/hq ADR 0219).
+func TestAPlanIsAnsweredOnlyByTheBuildItAskedFor(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	asked := time.Now().UTC().Add(-time.Minute)
+	plan := inventory.Plan{ID: "plan-own", Repository: "novox/a", Commit: "c0ffee", Created: asked,
+		State: inventory.PlanBuilding, Tiers: [][]string{{"a"}},
+		Modules: map[string]*inventory.PlanModule{"a": {State: "asked", AskedAt: &asked, Build: "build-own"}}}
+	if err := open.inventory.SavePlan(ctx, plan); err != nil {
+		t.Fatal(err)
+	}
+	planBuilt(ctx, open, "a", "0ldc0mm1t", "", time.Now().UTC(), "build-replay")
+	p, _ := open.inventory.PlanByID(ctx, plan.ID)
+	if p.Modules["a"].State != "asked" {
+		t.Fatalf("a replay asked after the plan settled it: %+v", p.Modules["a"])
+	}
+	// From the records too: a later build of the module recorded is not the plan's.
+	recorded := map[string][]inventory.Build{"a": {{ID: "build-replay", Commit: "0ldc0mm1t", Asked: time.Now(), At: time.Now()}}}
+	if settleFromRecords(&p, p.Tiers[0], recorded, nil) {
+		t.Fatalf("the records settled it with another build: %+v", p.Modules["a"])
+	}
+	planBuilt(ctx, open, "a", "c0ffee", "", asked, "build-own")
+	if p, _ = open.inventory.PlanByID(ctx, plan.ID); p.Modules["a"].State != "built" || p.Modules["a"].Commit != "c0ffee" {
+		t.Fatalf("its own build did not settle it: %+v", p.Modules["a"])
+	}
+}
+
+// rebuild of a build made at a commit asks what the module follows now, never the commit.
+func TestARebuildOfACommitAsksWhatTheModuleFollows(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	asked := asksRecorded(t)
+	twoTiers(t, open)
+	var refs []string
+	was := askABuild
+	askABuild = func(c context.Context, source buildSource, path, ref string) (string, error) {
+		refs = append(refs, ref)
+		return was(c, source, path, ref)
+	}
+	if err := open.inventory.RecordBuild(ctx, inventory.Build{ID: "build-at-commit", Repository: "novox/a",
+		Ref: "0123456789abcdef0123456789abcdef01234567", Module: "a", Commit: "0123456789abcdef0123456789abcdef01234567"}); err != nil {
+		t.Fatal(err)
+	}
+	if err := rebuildCommand(ctx, []string{"build-at-commit"}); err != nil {
+		t.Fatal(err)
+	}
+	if len(refs) != 1 || refs[0] != "main" || len(*asked) != 1 {
+		t.Fatalf("asked %v at %v", *asked, refs)
+	}
+}
+
+// An ask the worker counts out that no machine said it started is cancelled only if no start comes
+// while a holder looks: one that does withdraws the cancel, and kill is what ends it.
+func TestCancelOfAnAskNobodySaidIsWithdrawnWhenItStarts(t *testing.T) {
+	js, ask := aBuildQueue(t)
+	open := aMesh(t)
+	ctx := t.Context()
+	was := holderLooks
+	holderLooks = 300 * time.Millisecond
+	t.Cleanup(func() { holderLooks = was })
+	id := link.NewBuildID(time.Now())
+	ask(id, "https://forge.example/novox/a.git")
+	worker, _ := broker.HolderConsumerFor("", "", broker.DeclaredSeat{Name: link.TheBuildMachine, Accepts: []string{"build"}})
+	sub, err := js.Context().PullSubscribe(worker.Filters[0], worker.Name, nats.Bind(worker.Stream, worker.Name), nats.ManualAck())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = sub.Unsubscribe() }()
+	if _, err := sub.Fetch(1, nats.MaxWait(3*time.Second)); err != nil {
+		t.Fatal(err)
+	}
+	q, err := link.ReadQueue(ctx, js, link.TheBuildMachine)
+	if err != nil {
+		t.Fatal(err)
+	}
+	a, _ := q.Find(id)
+	if a.State != link.AskInFlight || a.On != "" {
+		t.Fatalf("the taken ask reads %+v", a)
+	}
+	// The holder that took it says it started, while the cancel waits.
+	go func() {
+		time.Sleep(100 * time.Millisecond)
+		started, _ := json.Marshal(link.BuildStart{ID: id, On: "ace", At: time.Now().UTC().Format(time.RFC3339Nano)})
+		_, _ = js.Context().Publish(link.BuildStarted(), started)
+	}()
+	if _, err := cancelAsk(ctx, js, open, link.TheBuildMachine, a); err == nil || !strings.Contains(err.Error(), "started on ace") {
+		t.Fatalf("a build that started was cancelled: %v", err)
+	}
+	if cancelled, _ := link.IsCancelled(js.Conn(), link.TheBuildMachine, id); cancelled {
+		t.Error("the withdrawn cancel is still marked")
+	}
+	if _, found, _ := open.inventory.BuildByID(ctx, id); found {
+		t.Error("a withdrawn cancel recorded an outcome")
+	}
+}

cmd/mesh-controller/readable.go

@@ -0,0 +1,274 @@
+package main
+
+import (
+	"encoding/json"
+	"fmt"
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"sort"
+	"time"
+)
+
+// The same answers, in a shape something other than a person can read.
+//
+// A board reads through interfaces and holds nothing (novox/hq 03-DESIGN/01-to-be/11-a-board.md).
+// Everything it needs is already answered by these commands — as text, for people, which is not
+// something a page can read. So each of them can say it again as JSON.
+//
+// **`--json` rather than a serving API**, because nothing needs one yet: whatever serves a board
+// runs the command, and the constraint in the design holds either way — the board never touches a
+// context's store. An API is the larger thing and should wait until something is asking for it.
+//
+// **These shapes are hard to change once anything is built against them.** So they stay close to
+// what the domain already calls things, and carry no summary field that would have to be kept
+// true. Nothing here is derived that a reader could not derive.
+
+// meshStatus is what `status --json` says: the three questions, in the order they are asked.
+type meshStatus struct {
+	// Wrong is every machine whose last declaration was refused or partly failed.
+	Wrong []machineDoing `json:"wrong"`
+	// Quiet is every machine not heard from lately. Not the same as wrong: new, switched off and
+	// unreachable are not "tried and could not".
+	Quiet []machineQuiet `json:"quiet"`
+	// Behind is every module built from something older than its source has.
+	Behind []moduleBehind `json:"behind"`
+	// Waiting is every machine not running what the mesh would send it. The same question as
+	// Behind one level down: that says the catalogue is old, this says a machine is — and only
+	// this one has somebody's change waiting inside it.
+	Waiting []machineWaiting `json:"waiting"`
+	// Reported is every machine's last word beside when it was last sent a declaration. A
+	// machine whose report is newer than its send has acted on the current declaration; one
+	// whose is older is still working — and Waiting cannot tell those apart, because the sent
+	// digest is recorded at send, not at apply.
+	Reported []machineReported `json:"reported"`
+	// Plans is what the last merges produced and where each stands (novox/hq ADR 0162): the
+	// open ones first, each saying its tier, what it waits for, and whether it has waited too long.
+	Plans []planStatus `json:"plans"`
+	// Unresolved is every machine that cannot be worked out at all, with what the mesh said when
+	// it tried. **A machine here is in none of the lists above**: nothing was computed for it, so
+	// there is nothing to compare it against and nothing it can be behind — which is why a
+	// document without this field described a wholly blocked mesh as a well one.
+	//
+	// Per machine, and data. One node failing must never take the document away from a reader
+	// asking about the others.
+	Unresolved []machineUnresolved `json:"unresolved"`
+	// Network is why the private network could not be computed, when it could not; absent when it
+	// could. Almost always a consequence of Unresolved: a node that does not resolve is not on the
+	// network, and a mesh whose hub is that node has no hub.
+	Network string `json:"network,omitempty"`
+	// Machines is how many the mesh knows about, so a reader can tell "none wrong" from
+	// "none at all".
+	Machines int `json:"machines"`
+	// Adopted is every node still adopted (novox/hq ADR 0100); absent when none is.
+	Adopted []string `json:"adopted,omitempty"`
+	// Untaken is every module assigned to a machine that is holding what it found rather than
+	// running what the module declares, because nothing took it (novox/hq 04-ISSUES/125). Absent
+	// when nothing is held.
+	//
+	// **A document without this said an outage was a well mesh.** Read from what each machine
+	// reported, so it is the machine's account and not the mesh's take-time listing.
+	Untaken []machineUntaken `json:"untaken,omitempty"`
+	// Filtered is every converged machine that is not filtered by the mesh alone (novox/hq ADR
+	// 0168), one entry per rule set the mesh did not write — the found firewall in force again,
+	// or a chain nobody speaks for. Absent when every converged machine is filtered by the mesh
+	// alone. A document without this called a machine well while a predecessor's chain refused
+	// what the mesh declared open.
+	Filtered []machineFiltered `json:"filtered,omitempty"`
+	// Unheld is every module on a machine whose resources are applied through a seat nothing on
+	// that machine holds, with the modules that could hold it (novox/hq ADR 0207). Absent when every
+	// dependency is met. Reported, not refused, until the switch.
+	Unheld []catalogue.Unheld `json:"unheld,omitempty"`
+	// Failing is every consumer a provider says it keeps failing, with the class of error, since
+	// when, and when it was last said (novox/hq ADR 0224). Absent when no provider says so. A
+	// document without this called the mesh well while the identity provider refused every consumer
+	// for a day (04-ISSUES/179).
+	Failing []inventory.ProviderStanding `json:"failing,omitempty"`
+}
+
+// machineFiltered is one rule set on a converged machine that the mesh did not write and that
+// refuses traffic: where it is, whose the host reads it as, and what it refuses.
+type machineFiltered struct {
+	Node    string `json:"node"`
+	Where   string `json:"where"`
+	Owner   string `json:"owner"`
+	Refuses string `json:"refuses"`
+}
+
+// machineUntaken is one module a machine is holding rather than running, and how many resources of
+// it are held.
+type machineUntaken struct {
+	Node   string `json:"node"`
+	Module string `json:"module"`
+	// Held is how many of the module's resources the machine is keeping as it found them. Zero is
+	// impossible here: a module with nothing held is not in this list.
+	Held int `json:"held"`
+}
+
+type machineUnresolved struct {
+	Node string `json:"node"`
+	// Problem is the mesh's own words, whole — newlines and all. It lists every requirement that
+	// could not be met, and a first line alone would name one of them and hide the rest.
+	Problem string `json:"problem"`
+}
+
+type machineDoing struct {
+	Node string `json:"node"`
+	// Outcome is refused or failed. Kept distinct all the way out: they are fixed in different
+	// places, and one word for both sends half the readers to the wrong one.
+	Outcome string `json:"outcome"`
+	Refused string `json:"refused,omitempty"`
+	Failed  []struct {
+		ID    string `json:"id"`
+		Error string `json:"error"`
+	} `json:"failed,omitempty"`
+	Applied int       `json:"applied"`
+	At      time.Time `json:"at"`
+	// Since is when this same failure was first reported and Times how many reports in a row
+	// have said it; Stuck is the mesh's word for "enough of them" (novox/hq 04-ISSUES/065).
+	Since *time.Time `json:"since,omitempty"`
+	Times int        `json:"times"`
+	Stuck bool       `json:"stuck"`
+}
+
+type machineReported struct {
+	Node    string     `json:"node"`
+	Outcome string     `json:"outcome"`
+	At      *time.Time `json:"at,omitempty"`
+	Sent    *time.Time `json:"sent,omitempty"`
+	// Current is whether the last report names the declaration last sent. Not derivable from
+	// the timestamps beside it: an apply begun under the previous declaration reports after the
+	// next send, newer and still about the old words.
+	Current bool `json:"current"`
+}
+
+type machineWaiting struct {
+	Node string `json:"node"`
+	// Never is true when nothing has ever been sent to it. Not out of date: nobody has ever asked
+	// this machine to be anything, and the two read differently to whoever is looking.
+	Never bool       `json:"never"`
+	Sent  *time.Time `json:"sent,omitempty"`
+}
+
+type machineQuiet struct {
+	Node string `json:"node"`
+	// LastSeen is absent when the machine has never spoken, which is a different thing from
+	// having been quiet for a while.
+	LastSeen *time.Time `json:"lastSeen,omitempty"`
+}
+
+type moduleBehind struct {
+	Module    string   `json:"module"`
+	BuiltFrom string   `json:"builtFrom"`
+	Head      string   `json:"head"`
+	On        []string `json:"on"`
+}
+
+// statusAsJSON answers the same questions as the text form, from the same reading.
+//
+// **It takes the whole reading rather than a growing argument list**, which is what let a new
+// answer be added to the text form and forgotten here — the two are one function's output in two
+// shapes, and they must not be able to differ about what was asked.
+//
+// It never fails on account of the mesh. Every per-machine problem in here is a field, so one
+// machine that cannot be worked out cannot stop a caller reading about the others: a
+// machine-readable interface that stops being machine-readable exactly when something is wrong is
+// one nobody can build an alarm on.
+func statusAsJSON(asked answers) ([]byte, error) {
+	wrong, nodes, quiet := asked.wrong, asked.nodes, asked.quiet
+	behind, sources := asked.behind, asked.sources
+	waiting, reported := asked.waiting, asked.reported
+
+	out := meshStatus{Machines: len(nodes), Wrong: []machineDoing{},
+		Quiet: []machineQuiet{}, Behind: []moduleBehind{}, Waiting: []machineWaiting{},
+		Reported: []machineReported{}, Unresolved: []machineUnresolved{},
+		Network: asked.network, Adopted: adoptedNodes(nodes), Plans: planStatuses(asked.plans, time.Now(), asked.paused)}
+	// In a stated order, so two readings of an unchanged mesh are the same document.
+	untakenNodes := make([]string, 0, len(asked.untaken))
+	for name := range asked.untaken {
+		untakenNodes = append(untakenNodes, name)
+	}
+	sort.Strings(untakenNodes)
+	for _, name := range untakenNodes {
+		modules := make([]string, 0, len(asked.untaken[name]))
+		for m := range asked.untaken[name] {
+			modules = append(modules, m)
+		}
+		sort.Strings(modules)
+		for _, m := range modules {
+			out.Untaken = append(out.Untaken,
+				machineUntaken{Node: name, Module: m, Held: asked.untaken[name][m]})
+		}
+	}
+	filteredNodes := make([]string, 0, len(asked.filtered))
+	for name := range asked.filtered {
+		filteredNodes = append(filteredNodes, name)
+	}
+	sort.Strings(filteredNodes)
+	for _, name := range filteredNodes {
+		f := asked.filtered[name]
+		if fw := f.FoundFirewall; fw != nil && fw.Active {
+			out.Filtered = append(out.Filtered, machineFiltered{Node: name, Where: "the found firewall",
+				Owner: inventory.FilterFoundFirewall, Refuses: fw.Kind + " is in force again"})
+		}
+		for _, x := range f.Others() {
+			out.Filtered = append(out.Filtered, machineFiltered{Node: name, Where: x.Where, Owner: x.Owner, Refuses: x.Refuses})
+		}
+	}
+	out.Unheld = asked.unheld
+	out.Failing = asked.failing
+	for name := range asked.refused {
+		out.Unresolved = append(out.Unresolved, machineUnresolved{
+			Node: name, Problem: asked.refused[name]})
+	}
+	sort.Slice(out.Unresolved, func(i, j int) bool {
+		return out.Unresolved[i].Node < out.Unresolved[j].Node
+	})
+	for _, r := range reported {
+		out.Reported = append(out.Reported, machineReported{
+			Node: r.Node, Outcome: r.Outcome, At: r.At, Sent: r.Sent, Current: r.Current})
+	}
+
+	for _, m := range waiting {
+		out.Waiting = append(out.Waiting, machineWaiting{
+			Node: m.Node, Never: m.Never, Sent: m.SentAt})
+	}
+
+	for _, d := range wrong {
+		row := machineDoing{
+			Node: d.Node, Outcome: d.Outcome, Refused: d.Refused, Applied: d.Applied, At: d.At,
+			Since: d.Since, Times: d.Times, Stuck: d.Stuck(),
+		}
+		for _, f := range d.Failed {
+			row.Failed = append(row.Failed, struct {
+				ID    string `json:"id"`
+				Error string `json:"error"`
+			}{ID: f.ID, Error: f.Error})
+		}
+		out.Wrong = append(out.Wrong, row)
+	}
+	for _, n := range quiet {
+		row := machineQuiet{Node: n.Name}
+		if !n.LastSeen.IsZero() {
+			seen := n.LastSeen
+			row.LastSeen = &seen
+		}
+		out.Quiet = append(out.Quiet, row)
+	}
+	for module, on := range behind {
+		from := sources[module]
+		out.Behind = append(out.Behind, moduleBehind{
+			Module: module, BuiltFrom: from.BuiltFrom, Head: from.Head, On: on,
+		})
+	}
+	return json.MarshalIndent(out, "", "  ")
+}
+
+// say prints a value as JSON, for the commands that can answer either way.
+func say(value any) error {
+	body, err := json.MarshalIndent(value, "", "  ")
+	if err != nil {
+		return err
+	}
+	fmt.Println(string(body))
+	return nil
+}

cmd/mesh-controller/readable_test.go

@@ -0,0 +1,212 @@
+package main
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// The shape something other than a person reads.
+//
+// Hard to change once anything is built against it, so it stays close to what the domain already
+// calls things and carries no summary field that would have to be kept true.
+
+func statusOf(t *testing.T, wrong []inventory.Doing, nodes, quiet []inventory.Node,
+	behind map[string][]string, sources map[string]inventory.Source) map[string]any {
+	t.Helper()
+	body, err := statusAsJSON(answers{
+		wrong: wrong, nodes: nodes, quiet: quiet, behind: behind, sources: sources})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var parsed map[string]any
+	if err := json.Unmarshal(body, &parsed); err != nil {
+		t.Fatalf("what a board would read is not JSON: %v", err)
+	}
+	return parsed
+}
+
+func TestRefusedAndFailedStayDistinctAllTheWayOut(t *testing.T) {
+	// They are fixed in different places, so one word for both would send half the readers of a
+	// page to the wrong one.
+	got := statusOf(t,
+		[]inventory.Doing{
+			{Node: "one", Outcome: inventory.OutcomeRefused, Refused: "a file needs a path"},
+			{Node: "two", Outcome: inventory.OutcomeFailed, Applied: 3,
+				Failed: []inventory.FailedResource{{ID: "shell.pkg", Error: "target not found"}}},
+		},
+		[]inventory.Node{{Name: "one"}, {Name: "two"}}, nil, nil, nil)
+
+	wrong, _ := got["wrong"].([]any)
+	if len(wrong) != 2 {
+		t.Fatalf("got %v", got["wrong"])
+	}
+	first, _ := wrong[0].(map[string]any)
+	if first["outcome"] != inventory.OutcomeRefused || first["refused"] == nil {
+		t.Fatalf("a refusal did not survive: %v", first)
+	}
+	second, _ := wrong[1].(map[string]any)
+	if second["outcome"] != inventory.OutcomeFailed {
+		t.Fatalf("a failure did not survive: %v", second)
+	}
+	if second["applied"] != float64(3) {
+		// "Three of eight" and "none of eight" are different machines.
+		t.Fatalf("what did apply was not carried: %v", second)
+	}
+}
+
+func TestAMachineThatNeverSpokeSaysSoByOmission(t *testing.T) {
+	// Never heard from and quiet for a while are different situations, and a zero time would read
+	// as a date in 1970 on any page that formatted it.
+	got := statusOf(t, nil,
+		[]inventory.Node{{Name: "silent"}, {Name: "away", LastSeen: time.Now().Add(-3 * time.Hour)}},
+		[]inventory.Node{{Name: "silent"}, {Name: "away", LastSeen: time.Now().Add(-3 * time.Hour)}},
+		nil, nil)
+
+	quiet, _ := got["quiet"].([]any)
+	if len(quiet) != 2 {
+		t.Fatalf("got %v", got["quiet"])
+	}
+	never, _ := quiet[0].(map[string]any)
+	if _, said := never["lastSeen"]; said {
+		t.Fatalf("a machine that never spoke carries a time: %v", never)
+	}
+	away, _ := quiet[1].(map[string]any)
+	if _, said := away["lastSeen"]; !said {
+		t.Fatalf("a machine that has been quiet carries no time: %v", away)
+	}
+}
+
+func TestNothingWrongIsAnEmptyListRatherThanNothing(t *testing.T) {
+	// A page distinguishing "no machines are wrong" from "this field is missing" would have to
+	// handle both, and null is the one that gets forgotten.
+	got := statusOf(t, nil, []inventory.Node{{Name: "a", LastSeen: time.Now()}}, nil, nil, nil)
+	for _, key := range []string{"wrong", "quiet", "behind"} {
+		list, ok := got[key].([]any)
+		if !ok {
+			t.Fatalf("%q is %T, not a list", key, got[key])
+		}
+		if len(list) != 0 {
+			t.Fatalf("%q is not empty: %v", key, list)
+		}
+	}
+	// And how many machines there are, so a reader can tell "none wrong" from "none at all".
+	if got["machines"] != float64(1) {
+		t.Fatalf("got %v", got["machines"])
+	}
+}
+
+func TestWhatIsBehindNamesTheMachinesRunningTheOldOne(t *testing.T) {
+	// A module being out of date is a fact about the catalogue; machines running the old one is
+	// the thing with consequences.
+	got := statusOf(t, nil, nil, nil,
+		map[string][]string{"shell": {"workstation", "laptop"}},
+		map[string]inventory.Source{"shell": {BuiltFrom: "aaaaaaa1", Head: "bbbbbbb2"}})
+
+	behind, _ := got["behind"].([]any)
+	if len(behind) != 1 {
+		t.Fatalf("got %v", got["behind"])
+	}
+	row, _ := behind[0].(map[string]any)
+	if row["module"] != "shell" || row["builtFrom"] != "aaaaaaa1" || row["head"] != "bbbbbbb2" {
+		t.Fatalf("got %v", row)
+	}
+	on, _ := row["on"].([]any)
+	if len(on) != 2 {
+		t.Fatalf("the machines running the old one are not named: %v", row)
+	}
+}
+
+func TestNoSecretIsInWhatABoardReads(t *testing.T) {
+	// Everything here comes from the mesh's own records, which hold no readable secret — but a
+	// shape a page is built against is exactly where one would eventually be added for
+	// convenience, so this says it out loud.
+	body, err := statusAsJSON(answers{
+		wrong: []inventory.Doing{{Node: "a", Outcome: inventory.OutcomeRefused,
+			Refused: "resource \"x\": a file needs a path"}},
+		nodes:   []inventory.Node{{Name: "a"}},
+		refused: map[string]string{"a": "nothing provides \"database\", wanted by web"},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, word := range []string{"password", "secret", "sealed", "credential", "token"} {
+		if strings.Contains(strings.ToLower(string(body)), word) {
+			t.Fatalf("what a board reads carries a %q field:\n%s", word, body)
+		}
+	}
+}
+
+// A failure that has been reported identically enough times is said to be stuck, with when it
+// began and how many times — so a board can tell a machine looping on something that will never
+// apply from one that failed a minute ago (novox/hq 04-ISSUES/065, ADR 0090).
+func TestAMachineFailingTheSameWayIsSaidToBeStuck(t *testing.T) {
+	began := time.Date(2026, 9, 21, 9, 0, 0, 0, time.UTC)
+	got := statusOf(t,
+		[]inventory.Doing{
+			{Node: "looping", Outcome: inventory.OutcomeFailed, Since: &began, Times: inventory.StuckAfter,
+				Failed: []inventory.FailedResource{{ID: "img", Error: "no such image"}}},
+			{Node: "once", Outcome: inventory.OutcomeFailed, Since: &began, Times: 1,
+				Failed: []inventory.FailedResource{{ID: "img", Error: "no such image"}}},
+		},
+		[]inventory.Node{{Name: "looping"}, {Name: "once"}}, nil, nil, nil)
+
+	wrong, _ := got["wrong"].([]any)
+	looping, _ := wrong[0].(map[string]any)
+	if looping["stuck"] != true || looping["times"] != float64(inventory.StuckAfter) {
+		t.Fatalf("three identical failures are stuck: %v", looping)
+	}
+	if since, _ := looping["since"].(string); !strings.HasPrefix(since, "2026-09-21T09:00:00") {
+		t.Fatalf("when it began was not carried: %v", looping)
+	}
+	once, _ := wrong[1].(map[string]any)
+	if once["stuck"] != false || once["times"] != float64(1) {
+		t.Fatalf("one failure is not stuck: %v", once)
+	}
+}
+
+// A converged machine something other than the mesh filters is named, per rule set, and is not
+// well (novox/hq ADR 0168); one filtered by the mesh alone is not in the list.
+func TestAMachineNotFilteredByTheMeshAloneIsNamedAndNotWell(t *testing.T) {
+	alone := inventory.Filtering{Filters: []inventory.Filter{
+		{Where: "table inet mesh, chain input", Owner: inventory.FilterMesh, Refuses: "policy drop"},
+		{Where: "table ip filter, chain DOCKER", Owner: inventory.FilterRuntime, Refuses: "drop"},
+		{Where: "table ip filter, chain f2b-sshd", Owner: inventory.FilterBan, Refuses: "ip saddr 192.0.2.1 reject"},
+	}}
+	if !alone.Alone() {
+		t.Fatal("the mesh's own, the runtime's and a ban are not the mesh alone")
+	}
+	notAlone := inventory.Filtering{
+		Filters: append(alone.Filters, inventory.Filter{Where: "chain HAL-MESH-ONLY (iptables-legacy)",
+			Owner: inventory.FilterOther, Refuses: `-A HAL-MESH-ONLY -m comment --comment "not public" -j DROP`}),
+		FoundFirewall: &inventory.FoundFirewall{Kind: "ufw", Active: true},
+	}
+	asked := answers{nodes: []inventory.Node{{Name: "home-server"}, {Name: "laptop"}},
+		filtered: map[string]inventory.Filtering{"home-server": notAlone}}
+	if asked.well() {
+		t.Fatal("a machine not filtered by the mesh alone reads as well")
+	}
+	body, err := statusAsJSON(asked)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var parsed struct {
+		Filtered []map[string]string `json:"filtered"`
+	}
+	if err := json.Unmarshal(body, &parsed); err != nil {
+		t.Fatal(err)
+	}
+	if len(parsed.Filtered) != 2 {
+		t.Fatalf("filtered: %v", parsed.Filtered)
+	}
+	if parsed.Filtered[0]["node"] != "home-server" || parsed.Filtered[0]["owner"] != inventory.FilterFoundFirewall ||
+		parsed.Filtered[1]["where"] != "chain HAL-MESH-ONLY (iptables-legacy)" || parsed.Filtered[1]["owner"] != inventory.FilterOther {
+		t.Fatalf("filtered: %v", parsed.Filtered)
+	}
+	if body, _ := statusAsJSON(answers{nodes: asked.nodes}); strings.Contains(string(body), `"filtered"`) {
+		t.Fatal("a mesh filtered by itself alone carries a filtered list")
+	}
+}

cmd/mesh-controller/references.go

@@ -0,0 +1,150 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// What a build is recorded as, and what it is announced and handed on as.
+//
+// **Recorded by what it is; routed where it is used** (novox/hq 04-ISSUES/102). The builder
+// announces each artifact by the reference it pushed — `<registry>:<port>/<module>/<artifact>@sha256:…`
+// — and the manifest with those references in it. The mesh records the digest and the path
+// (catalogue.Recorded) and composes the store's address back in wherever a machine or a builder
+// needs a reference it can fetch: a declaration, a replay to the catalogue, the bases a build is
+// given. Nothing recorded carries a port, so moving the store is a settings change and not a
+// rebuild of everything the mesh has ever built.
+
+// recordedManifest is a build's manifest with the store's address taken off every reference the
+// build itself made. Other references — an image a module runs from a public registry — are what
+// they were, which is why this rewrites only what `made` names rather than everything that looks
+// like an address.
+func recordedManifest(raw json.RawMessage, made []inventory.Artifact) json.RawMessage {
+	announced := map[string]bool{}
+	for _, a := range made {
+		announced[a.Reference] = true
+	}
+	return withReferences(raw, func(ref string) (string, bool) {
+		if !announced[ref] {
+			return ref, false
+		}
+		return catalogue.Recorded(ref), true
+	})
+}
+
+// routedManifest is a recorded manifest with the store's address, as this network reaches it now,
+// composed into every reference the build made — recorded either way, before or after references
+// were kept without their address.
+func routedManifest(raw json.RawMessage, made []inventory.Artifact, address string) json.RawMessage {
+	recorded := map[string]bool{}
+	for _, a := range made {
+		recorded[catalogue.Recorded(a.Reference)] = true
+	}
+	return withReferences(raw, func(ref string) (string, bool) {
+		if !recorded[catalogue.Recorded(ref)] {
+			return ref, false
+		}
+		return catalogue.Rerouted(ref, address), true
+	})
+}
+
+// withReferences applies `rewrite` to each resource's `image` and `source`, and hands the manifest
+// back untouched — byte for byte — when nothing changed or it could not be read: what a manifest
+// is, is the parser's to say, and it says so with a better sentence than anything here would.
+func withReferences(raw json.RawMessage, rewrite func(string) (string, bool)) json.RawMessage {
+	if len(raw) == 0 {
+		return raw
+	}
+	var manifest map[string]any
+	if err := json.Unmarshal(raw, &manifest); err != nil {
+		return raw
+	}
+	resources, _ := manifest["resources"].([]any)
+	changed := false
+	for _, r := range resources {
+		resource, ok := r.(map[string]any)
+		if !ok {
+			continue
+		}
+		for _, key := range []string{"image", "source"} {
+			if written, ok := resource[key].(string); ok {
+				if rewritten, did := rewrite(written); did && rewritten != written {
+					resource[key] = rewritten
+					changed = true
+				}
+			}
+		}
+	}
+	if !changed {
+		return raw
+	}
+	out, err := json.Marshal(manifest)
+	if err != nil {
+		return raw
+	}
+	return out
+}
+
+// routedArtifacts is a build's artifacts as something can fetch them now.
+func routedArtifacts(made []inventory.Artifact, address string) []inventory.Artifact {
+	if address == "" {
+		return made
+	}
+	out := make([]inventory.Artifact, 0, len(made))
+	for _, a := range made {
+		out = append(out, inventory.Artifact{Name: a.Name, Kind: a.Kind,
+			Reference: catalogue.Rerouted(a.Reference, address)})
+	}
+	return out
+}
+
+// whereTheStoreIs is the artifact store's address as something on `forNode` reaches it, or "" —
+// read for a caller that has the inventory open and nothing else in hand. With no node named, the
+// store's own node: loopback when nothing is on the network yet.
+func whereTheStoreIs(ctx context.Context, inv *inventory.Inventory, forNode string) (string, error) {
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return "", err
+	}
+	if forNode == "" {
+		if holder, _, found, err := artifactStoreHolder(ctx, inv); err != nil {
+			return "", err
+		} else if found {
+			forNode = holder
+		}
+	}
+	return artifactStoreAddress(ctx, inv, shelf, forNode)
+}
+
+// whereABuilderReachesTheStore is the store's address for the machine that builds: the network's
+// when there is one, else loopback on the store's own node — when that node also holds a module
+// requiring the store, which is what a builder is (genesis: one node holds both).
+func whereABuilderReachesTheStore(ctx context.Context, inv *inventory.Inventory) (string, error) {
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return "", err
+	}
+	holder, _, found, err := artifactStoreHolder(ctx, inv)
+	if err != nil || !found {
+		return "", err
+	}
+	assigned, err := inv.Assigned(ctx, holder)
+	if err != nil {
+		return "", err
+	}
+	besideIt := false
+	for _, a := range assigned {
+		for _, r := range shelf[a].Requires {
+			if r == catalogue.ArtifactStoreProvision {
+				besideIt = true
+			}
+		}
+	}
+	if !besideIt {
+		holder = ""
+	}
+	return artifactStoreAddress(ctx, inv, shelf, holder)
+}

cmd/mesh-controller/release_plan_test.go

@@ -0,0 +1,214 @@
+package main
+
+import (
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A merge produces a tiered plan (novox/hq ADR 0162): what moved and everything reachable from it,
+// sorted so a tier depends only on earlier ones — with the three kinds of dependency told apart.
+func TestAMergeIsPlannedInTiersAlongTheThreeKindsOfDependency(t *testing.T) {
+	edges := []inventory.Edge{
+		// build dependencies: images on the runtime, a plugin on one of them
+		{From: "shop", To: "mesh-tools", Kind: inventory.EdgeStandsOn},
+		{From: "postgres", To: "mesh-tools", Kind: inventory.EdgeStandsOn},
+		{From: "shop-plugin", To: "shop", Kind: inventory.EdgeDeclared},
+		// a code dependency: the proxy packages the controller's source — same tier
+		{From: "route-proxy", To: "mesh-controller", Kind: inventory.EdgePackages},
+		{From: "builder", To: "mesh-controller", Kind: inventory.EdgePackages},
+		// runtime dependencies: everything source-built is built by the builder
+		{From: "shop", To: "builder", Kind: inventory.EdgeBuiltBy},
+		{From: "postgres", To: "builder", Kind: inventory.EdgeBuiltBy},
+		{From: "shop-plugin", To: "builder", Kind: inventory.EdgeBuiltBy},
+		{From: "route-proxy", To: "builder", Kind: inventory.EdgeBuiltBy},
+		{From: "mesh-controller", To: "builder", Kind: inventory.EdgeBuiltBy},
+		{From: "mesh-tools", To: "builder", Kind: inventory.EdgeBuiltBy},
+		{From: "builder", To: "mesh-tools", Kind: inventory.EdgeStandsOn},
+		{From: "unrelated", To: "alpine", Kind: inventory.EdgeStandsOn},
+	}
+	// The runtime image moved: everything on it, and what is built by what is on it.
+	set := reachableFrom([]string{"mesh-tools"}, edges)
+	// What stands on the runtime, and the builder that stands on it; not the controller, which the
+	// builder merely builds, nor the proxy that packages the controller.
+	want := []string{"builder", "mesh-tools", "postgres", "shop", "shop-plugin"}
+	if len(set) != len(want) {
+		t.Fatalf("reachable from the runtime: %v, want %v", set, want)
+	}
+	tiers := tiersOf(set, edges)
+	pos := map[string]int{}
+	for i, tier := range tiers {
+		for _, m := range tier {
+			pos[m] = i
+		}
+	}
+	if pos["mesh-tools"] != 0 || pos["builder"] != 1 {
+		t.Fatalf("the runtime then the builder: %v", tiers)
+	}
+	if !(pos["shop"] > pos["builder"] && pos["postgres"] > pos["builder"]) {
+		t.Fatalf("what the builder builds comes after the builder: %v", tiers)
+	}
+	if pos["shop-plugin"] <= pos["shop"] {
+		t.Fatalf("a plugin after what it is declared on: %v", tiers)
+	}
+	if hasCycle(tiers, edges) {
+		t.Fatalf("no cycle here: %v", tiers)
+	}
+
+	// The controller alone moved: the proxy with it, nothing else.
+	small := reachableFrom([]string{"mesh-controller"}, edges)
+	if len(small) != 3 {
+		t.Fatalf("a controller merge rebuilds the controller and what packages it: %v", small)
+	}
+	// The builder packages the controller's source (same tier by that edge) and the controller is
+	// built by the builder (next tier by that one): the builder first, then the controller and the
+	// proxy together — a code dependency in one tier, a runtime dependency across tiers.
+	smallTiers := tiersOf(small, edges)
+	if len(smallTiers) != 2 || smallTiers[0][0] != "builder" || len(smallTiers[1]) != 2 {
+		t.Fatalf("the builder, then the controller and the proxy together: %v", smallTiers)
+	}
+	// The builder alone moved: the builder, and nothing it builds.
+	if only := reachableFrom([]string{"builder"}, edges); len(only) != 1 {
+		t.Fatalf("a build machine change rebuilds the build machine alone: %v", only)
+	}
+
+	// Only a runtime dependency gates on deployment, and only when the module rolls out.
+	p := planOfMerge(link.SourceMoved{Owner: "novox", Repo: "mesh-tools", Commit: "abc"}, []string{"mesh-tools"}, edges)
+	p.Tier = pos["builder"]
+	rollsOut := func(m string) bool { return m == "builder" }
+	if g := gates(p, edges, rollsOut); len(g) != 1 || g[0] != "builder" {
+		t.Fatalf("the builder gates the tier after it: %v", g)
+	}
+	p.Tier = 0
+	if g := gates(p, edges, rollsOut); len(g) != 0 {
+		t.Fatalf("the runtime image is a build dependency and gates nothing: %v", g)
+	}
+	if g := gates(p, edges, func(string) bool { return false }); len(g) != 0 {
+		t.Fatalf("a module that only records its upgrade gates nothing: %v", g)
+	}
+}
+
+// A gate is open once every machine running the module has reported after it was built.
+func TestAGateOpensWhenTheMachinesHaveReportedSinceTheBuild(t *testing.T) {
+	built := time.Date(2026, 10, 1, 15, 0, 0, 0, time.UTC)
+	before, after := built.Add(-time.Minute), built.Add(time.Minute)
+	reports := []inventory.Reported{{Node: "anchor", At: &after}, {Node: "home-server", At: &before}}
+	ok, waiting := applied("builder", built, []string{"anchor", "home-server"}, reports)
+	if ok || len(waiting) != 1 || waiting[0] != "home-server" {
+		t.Fatalf("one machine has not reported since the build: ok=%v waiting=%v", ok, waiting)
+	}
+	if ok, _ := applied("builder", built, []string{"anchor"}, reports); !ok {
+		t.Fatal("the machine that reported after the build holds the gate open")
+	}
+	if ok, _ := applied("builder", built, nil, reports); !ok {
+		t.Fatal("a module running nowhere gates nothing")
+	}
+}
+
+// A cycle is not lost: what remains is one last tier, and the caller says so.
+func TestACycleIsOneLastTierAndSaidSo(t *testing.T) {
+	edges := []inventory.Edge{{From: "a", To: "b", Kind: inventory.EdgeStandsOn}, {From: "b", To: "a", Kind: inventory.EdgeStandsOn}}
+	tiers := tiersOf([]string{"a", "b"}, edges)
+	if len(tiers) != 1 || len(tiers[0]) != 2 || !hasCycle(tiers, edges) {
+		t.Fatalf("a cycle should be one tier of two, said: %v", tiers)
+	}
+}
+
+// novox/hq 04-ISSUES/211: a merge moving the toolchain and a bundle compiled in it builds the
+// bundle a tier after the toolchain, not beside it.
+func TestABundleIsPlannedAfterTheToolchainItIsCompiledIn(t *testing.T) {
+	edges := []inventory.Edge{{From: "node-tools", To: "mesh-tools", Kind: inventory.EdgeStandsOn}}
+	p := planOfMerge(link.SourceMoved{Owner: "novox", Repo: "mesh-tools", Commit: "abc"},
+		[]string{"mesh-tools", "node-tools"}, edges)
+	if len(p.Tiers) != 2 || p.Tiers[0][0] != "mesh-tools" || p.Tiers[1][0] != "node-tools" {
+		t.Fatalf("the toolchain, then the bundle: %v", p.Tiers)
+	}
+}
+
+// novox/hq 04-ISSUES/214: a plan whose build outcome was recorded while no controller followed it —
+// the controller rebuilding itself — settles from the build records instead of waiting for ever.
+func TestAPlanSettlesAnAskedBuildFromTheRecords(t *testing.T) {
+	asked := time.Date(2026, 10, 3, 19, 20, 0, 0, time.UTC)
+	p := inventory.Plan{ID: "plan-1", Tiers: [][]string{{"mesh-controller", "builder"}, {"route-proxy"}},
+		Modules: map[string]*inventory.PlanModule{
+			"mesh-controller": {State: "asked", AskedAt: &asked},
+			"builder":         {State: "asked", AskedAt: &asked},
+		}}
+	records := map[string][]inventory.Build{
+		// Newest first, as Builds answers: the build after the ask is the outcome.
+		"mesh-controller": {
+			{ID: "build-2", Commit: "2ebbb799", At: asked.Add(4 * time.Minute)},
+			{ID: "build-1", Commit: "06ea2168", At: asked.Add(-10 * time.Minute)},
+		},
+		// Only a build from before the ask: not this ask's outcome.
+		"builder": {{ID: "build-0", Commit: "06ea2168", At: asked.Add(-time.Hour)}},
+	}
+	if !settleFromRecords(&p, p.Tiers[0], records, nil) {
+		t.Fatal("nothing settled, though the controller's build is recorded after the ask")
+	}
+	if s := p.Modules["mesh-controller"]; s.State != "built" || s.Commit != "2ebbb799" || s.BuiltAt == nil {
+		t.Errorf("the controller's ask is %+v, want built from 2ebbb799", s)
+	}
+	if s := p.Modules["builder"]; s.State != "asked" {
+		t.Errorf("an ask with no record after it was settled: %+v", s)
+	}
+
+	// novox/hq 04-ISSUES/219: a build recorded after the ask but asked before it — an earlier
+	// plan's late outcome — is not this ask's, built or failed.
+	r := inventory.Plan{ID: "plan-3", Tiers: [][]string{{"postgres"}},
+		Modules: map[string]*inventory.PlanModule{"postgres": {State: "asked", AskedAt: &asked}}}
+	late := map[string][]inventory.Build{"postgres": {
+		{ID: "build-old", Commit: "efff5415", Asked: asked.Add(-18 * time.Minute), At: asked.Add(12 * time.Minute)},
+	}}
+	if settleFromRecords(&r, r.Tiers[0], late, nil) || r.Modules["postgres"].State != "asked" {
+		t.Errorf("an earlier ask's late outcome settled this ask: %+v", r.Modules["postgres"])
+	}
+	// Newest heard first: the earlier ask's late outcome, then this ask's own, heard before it.
+	late["postgres"] = append(late["postgres"], inventory.Build{ID: "build-mine", Commit: "4bcd5f73",
+		Asked: asked.Add(time.Second), At: asked.Add(5 * time.Minute)})
+	if !settleFromRecords(&r, r.Tiers[0], late, nil) || r.Modules["postgres"].State != "built" ||
+		r.Modules["postgres"].Commit != "4bcd5f73" {
+		t.Errorf("this ask's own outcome, heard before the earlier ask's, did not settle it: %+v", r.Modules["postgres"])
+	}
+
+	// A failure recorded after the ask fails the plan, as hearing it would have.
+	q := inventory.Plan{ID: "plan-2", Tiers: [][]string{{"x"}},
+		Modules: map[string]*inventory.PlanModule{"x": {State: "asked", AskedAt: &asked}}}
+	settleFromRecords(&q, q.Tiers[0], map[string][]inventory.Build{"x": {{ID: "b", Failed: "no", At: asked.Add(time.Minute)}}}, nil)
+	if q.State != inventory.PlanFailed || q.Modules["x"].State != "failed" {
+		t.Errorf("a recorded failure did not fail the plan: %+v %+v", q, q.Modules["x"])
+	}
+}
+
+// The first machine's report, made between the send to it and the send to the rest, opens the gate for it:
+// each machine is judged from its own send, not from the last one (novox/hq issue 256).
+func TestTheGateJudgesEachMachineFromItsOwnSend(t *testing.T) {
+	built := time.Date(2026, 10, 5, 18, 22, 0, 0, time.UTC)
+	firstSent := built.Add(31 * time.Second)
+	firstReported := built.Add(43 * time.Second)
+	restSent := built.Add(58 * time.Second)
+	restReported := built.Add(74 * time.Second)
+	reports := []inventory.Reported{
+		{Node: "ace", At: &firstReported},
+		{Node: "g14", At: &restReported},
+	}
+	since := func(node string) time.Time {
+		if node == "ace" {
+			return firstSent
+		}
+		return restSent
+	}
+	if ok, waiting := appliedEach("build-agent", since, []string{"ace", "g14"}, reports); !ok {
+		t.Fatalf("the gate still waits on %v, though each reported after its own send", waiting)
+	}
+	// The old reading, every machine from the last send, is what held the plan.
+	if ok, _ := applied("build-agent", restSent, []string{"ace", "g14"}, reports); ok {
+		t.Fatal("the single-moment reading should hold the first machine back")
+	}
+	early := built.Add(10 * time.Second)
+	if ok, waiting := appliedEach("build-agent", since, []string{"ace"}, []inventory.Reported{{Node: "ace", At: &early}}); ok || waiting[0] != "ace" {
+		t.Fatal("a report from before the machine was sent opened the gate")
+	}
+}

cmd/mesh-controller/rollout.go

@@ -0,0 +1,457 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"strings"
+	"time"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/secrets"
+)
+
+// Moving the mesh's own traffic to the bus being built (novox/hq ADR 0116 step 5).
+//
+// **The whole mesh moves at once, so there is nothing to inspect afterwards.** Every seam ships both
+// transports and every one of them chooses by a single fact; this is the step that flips it. That
+// shape is deliberate — steps 1 to 4 leave every node where it is, so the cost of being wrong stays
+// bounded until here — and it means the useful work is almost all in the checking.
+//
+// So `rollout check` is the command that matters and the one that can be run any number of times
+// against a mesh that is serving. It answers from records: what is missing, and what would happen.
+// `rollout` itself refuses unless the check is clean.
+//
+// **The old broker goes with the move, and goes last** (novox/hq ADR 0131): AMQP is not a provision,
+// so once every machine reports on the new bus its module is unassigned. Only the mesh's own traffic
+// is what moves, which is why this is survivable at all: what breaks if it goes wrong is the mesh's
+// ability to change things, not the services its modules are serving — measured on 2026-09-27, when
+// a seat emptied mid-change and the control plane looped for two hours while every service stayed up.
+
+const rolloutUsage = "rollout check | rollout mint [--again] | rollout hand <node> | rollout --confirm"
+
+func rolloutCommand(ctx context.Context, args []string) error {
+	switch {
+	case len(args) == 1 && args[0] == "check":
+		return rolloutCheck(ctx)
+	case len(args) == 1 && args[0] == "mint":
+		return rolloutMint(ctx, false)
+	case len(args) == 2 && args[0] == "hand":
+		return rolloutHand(ctx, args[1])
+	case len(args) == 2 && args[0] == "mint" && args[1] == "--again":
+		// Every credential minted afresh, whether or not one exists — for a mint that was wrong
+		// before anything was pushed. Afterwards nothing that received the old one still works,
+		// which is fine exactly when nothing received it.
+		return rolloutMint(ctx, true)
+	case len(args) == 1 && args[0] == "--confirm":
+		return errors.New(
+			"the rollout itself is not built yet: `rollout check` answers whether it could run, and " +
+				"what is missing. Moving every node at once is the one step with nothing to inspect " +
+				"afterwards, so it is not being written before the check it depends on has been run " +
+				"against a real mesh")
+	default:
+		return errors.New(rolloutUsage)
+	}
+}
+
+// rolloutCheck says whether the mesh could move, and what would happen if it did.
+func rolloutCheck(ctx context.Context) error {
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	state, err := readinessOf(ctx, inv)
+	if err != nil {
+		return err
+	}
+
+	fmt.Println("the bus this mesh would move to")
+	if state.TheBus == "" {
+		fmt.Printf("  nothing names one (%s is unset)\n", broker.NATSVar)
+	} else {
+		standing := "not answering"
+		if state.ServerStanding {
+			standing = "answering"
+		}
+		fmt.Printf("  %s — %s\n", state.TheBus, standing)
+	}
+	fmt.Println()
+
+	fmt.Println("what would move")
+	for _, step := range broker.WhatMoves(state) {
+		fmt.Printf("  %s\n", step)
+	}
+	fmt.Println()
+
+	why := notReadyOf(state)
+	if len(why) == 0 {
+		fmt.Println("nothing is missing: this mesh could move its bus.")
+		fmt.Println()
+		fmt.Println("Read `what would move` above once more before running it. Every node moves at the")
+		fmt.Println("same moment and there is no half-moved state to look at afterwards.")
+		return nil
+	}
+	fmt.Printf("not ready — %d thing(s) to do first:\n", len(why))
+	for i, w := range why {
+		fmt.Printf("  %d. %s\n", i+1, w)
+	}
+	return nil
+}
+
+// readinessOf gathers what the mesh knows about its own ability to move.
+//
+// Reads and one dial, and nothing is written. Safe to run on a mesh that is serving, which is the
+// point: the answer is only useful if it can be had without committing to anything.
+func readinessOf(ctx context.Context, inv *inventory.Inventory) (broker.Readiness, error) {
+	state := broker.Readiness{
+		Credentialled:       map[string]bool{},
+		ModuleCredentialled: map[string]bool{},
+		// The old broker keeps its other clients on this installation, and saying so is how the plan
+		// stops reading as a retirement.
+	}
+
+	address, _, err := broker.OnNATS()
+	if err != nil {
+		return state, err
+	}
+	state.TheBus = address
+	if address != "" {
+		// One dial, briefly. "Is it answering" is the one fact records cannot hold, and a mesh about
+		// to move onto a server that is not there should hear it here rather than afterwards.
+		//
+		// **Dialled the way the mesh dials it** — credential and pin — because a bare connect to a
+		// bus that requires TLS and a user fails at the handshake, and the check then reported a
+		// standing server as absent (seen live, 2026-09-28).
+		if js, err := broker.Dial(address, nats.Timeout(5*time.Second)); err == nil {
+			state.ServerStanding = true
+			js.Close()
+		}
+	}
+
+	nodes, err := inv.Nodes(ctx)
+	if err != nil {
+		return state, err
+	}
+	kept, err := inv.BusUsers(ctx)
+	if err != nil {
+		return state, err
+	}
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return state, err
+	}
+
+	for _, n := range nodes {
+		state.Nodes = append(state.Nodes, n.Name)
+		_, has := kept[broker.Principal{Kind: broker.KindNode, Node: n.Name}.Username()]
+		state.Credentialled[n.Name] = has
+
+		assigned, err := inv.Assigned(ctx, n.Name)
+		if err != nil {
+			return state, err
+		}
+		for _, module := range assigned {
+			m, known := shelf[module]
+			if !known {
+				continue
+			}
+			// The machine that holds the bus seat is the one that would be sent the user list.
+			if m.BusUsers != "" && m.ClaimsSeat("mesh-broker") {
+				state.Holder = n.Name
+				state.AccountsComposed = wasSentTheUserList(ctx, inv, n.Name)
+			}
+			// A module that never speaks needs no credential, so it is not counted as missing one.
+			if !speaksOnTheBus(m) {
+				continue
+			}
+			named := n.Name + "/" + module
+			state.Modules = append(state.Modules, named)
+			_, hasOne := kept[broker.Principal{
+				Kind: broker.KindModule, Node: n.Name, Module: module,
+			}.Username()]
+			state.ModuleCredentialled[named] = hasOne
+		}
+	}
+	return state, nil
+}
+
+// speaksOnTheBus says whether a module reaches the bus at all.
+//
+// A third of the catalogue never does (novox/hq ADR 0120), and counting those as missing a credential
+// would bury the ones that matter under a list nobody can act on.
+func speaksOnTheBus(m catalogue.Manifest) bool {
+	return len(m.EmitsAll()) > 0 || len(m.Consumes) > 0 || len(m.Tools) > 0 ||
+		len(m.DefinesSeats) > 0 || len(m.Uses) > 0 || len(m.Claims) > 0
+}
+
+// wasSentTheUserList says whether the machine holding the bus has had a declaration since the user
+// list became part of one.
+//
+// Read from what the mesh recorded sending rather than asked of the machine: a machine that is away
+// has still been sent it, and this question is about whether the mesh did its part.
+func wasSentTheUserList(ctx context.Context, inv *inventory.Inventory, node string) bool {
+	digest, err := inv.Outstanding(ctx, node)
+	return err == nil && strings.TrimSpace(digest) != ""
+}
+
+// notReadyOf is the readiness reasoning, named here so a test can reach it without the command's
+// printing. The reasoning itself is the broker package's, where it is pure.
+func notReadyOf(state broker.Readiness) []string { return broker.NotReady(state) }
+
+// rolloutMint gives every principal the new bus will have a credential it does not yet have, and
+// puts each where its owner reads it (novox/hq design 28, task 5.2): a machine's as a membership
+// sealed into its declaration, a module's as its broker secret, the control plane's own as its
+// `bus` secret. Idempotent: what already has a hash is left alone, so running it again is harmless.
+//
+// **Before anything moves, and it is what makes moving possible.** A machine moved without a
+// credential cannot come back, and afterwards there is no bus to tell it anything over — which is
+// why `rollout check` refuses until this has run. The bus's address is worked out here, from where
+// the module that provides it is assigned, rather than read from this process's environment: this
+// process is still on the old bus when this runs, and must be.
+func rolloutMint(ctx context.Context, again bool) error {
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	known, err := broker.FromEnvironment()
+	if err != nil {
+		return fmt.Errorf("the bus's certificate is not known to this process, and every membership "+
+			"must carry its fingerprint: %w", err)
+	}
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return err
+	}
+	entries, err := inv.Catalogued(ctx)
+	if err != nil {
+		return err
+	}
+	var busNode, controllerNode string
+	for _, e := range entries {
+		switch {
+		case e.Manifest.ClaimsSeat("mesh-broker") && providesBus(e.Manifest) && len(e.On) > 0:
+			busNode = e.On[0]
+		case e.Manifest.Module == "mesh-controller" && len(e.On) > 0:
+			controllerNode = e.On[0]
+		}
+	}
+	if busNode == "" {
+		return errors.New("no assigned module provides mesh-bus and claims mesh-broker, so there is no " +
+			"bus to mint credentials for — register and assign it first")
+	}
+	onNetwork, err := whereEveryoneIs(ctx, inv, shelf)
+	if err != nil {
+		return err
+	}
+	busHost := onNetwork[busNode]
+	if busHost == "" {
+		// **The hub is not in that map.** The machine that took over the tunnel is where the current
+		// bus already answers, and every machine dials it at the address the mesh handed them — so
+		// when the new bus runs on the same machine, that address is the one to tell them, with the
+		// new port. Found live: the control node is the hub, and the map lists the machines placed
+		// around it.
+		// The host alone: no scheme (BareAddress adds one where none was, which is the wrong
+		// direction here — every URL built below adds its own) and no port.
+		_, _, host := broker.CredentialIn(known.Address)
+		if host == "" {
+			host = known.Address
+		}
+		if _, after, hasScheme := strings.Cut(host, "://"); hasScheme {
+			host = after
+		}
+		host = strings.TrimSpace(host)
+		if i := strings.LastIndex(host, ":"); i > 0 && !strings.Contains(host[i:], "]") {
+			host = host[:i]
+		}
+		if host == "" {
+			return fmt.Errorf("%s runs the new bus and has no address on the private network, and the "+
+				"current bus's address is unknown too, so no machine could be told where it is", busNode)
+		}
+		busHost = host
+	}
+	busAddress := busHost + ":4222"
+
+	records, err := inv.BusRecords(ctx)
+	if err != nil {
+		return err
+	}
+	users, err := broker.Users(records)
+	if err != nil {
+		return err
+	}
+	kept, err := inv.BusUsers(ctx)
+	if err != nil {
+		return err
+	}
+	hashes := make(map[string]string, len(kept))
+	for name, u := range kept {
+		hashes[name] = u.PasswordHash
+	}
+	_, missing := broker.WithPasswords(users, hashes)
+	wanted := map[string]bool{}
+	for _, m := range missing {
+		wanted[m] = true
+	}
+
+	var machines, modules, skipped int
+	for _, p := range users {
+		if !again && !wanted[p.Username()] {
+			continue
+		}
+		switch p.Kind {
+		case broker.KindController:
+			if controllerNode == "" {
+				return errors.New("the control plane is not assigned anywhere, so its credential has nowhere to go")
+			}
+			password, err := inv.MintBusPassword(ctx, inventory.BusUser{Username: p.Username(), Kind: inventory.BusController})
+			if err != nil {
+				return err
+			}
+			url := "nats://" + p.Username() + ":" + password + "@" + busAddress
+			if err := inv.AcceptSecretForModule(ctx, controllerNode, "mesh-controller", "bus", url); err != nil {
+				return fmt.Errorf("the control plane's credential is minted and could not be sealed to %s: %w", controllerNode, err)
+			}
+			fmt.Printf("control plane: credential minted, sealed to %s as its `bus` secret\n", controllerNode)
+
+		case broker.KindNode:
+			password, err := inv.MintBusPassword(ctx, inventory.BusUser{Username: p.Username(), Kind: inventory.BusNode, Node: p.Node})
+			if err != nil {
+				return err
+			}
+			membership, _ := json.Marshal(map[string]string{
+				"broker": busAddress, "fingerprint": known.Fingerprint, "password": password, "transport": "nats",
+			})
+			key, err := inv.SealingKeyOf(ctx, p.Node)
+			if err != nil {
+				return fmt.Errorf("%s has no sealing key, so its membership cannot be sealed to it: %w", p.Node, err)
+			}
+			sealed, err := secrets.Seal(key, membership)
+			if err != nil {
+				return err
+			}
+			if err := inv.PutBusMembership(ctx, p.Node, sealed); err != nil {
+				return err
+			}
+			machines++
+
+		case broker.KindModule, broker.KindNodeTools:
+			// The runtime is minted and delivered exactly as a module is (novox/hq ADR 0175): it is
+			// issued as the module it stands for, to that module's `broker` secret.
+			if p.Module == "mesh-controller" {
+				// The control plane is a module too, and its `broker` secret is the old bus's
+				// credential it is still using while this runs. Writing the new bus's blob there
+				// cut the mesh off from its own old bus mid-move (2026-09-28). Its new-bus credential
+				// is the controller principal's `bus` secret above; nothing else is needed here.
+				skipped++
+				continue
+			}
+			m, inShelf := shelf[p.Module]
+			if !inShelf {
+				skipped++
+				continue
+			}
+			if _, reads := m.OwnSecrets["broker"]; !reads {
+				fmt.Printf("  %s on %s speaks on the bus but declares no `broker` secret to receive a credential in; skipped\n", p.Module, p.Node)
+				skipped++
+				continue
+			}
+			password, err := inv.MintBusPassword(ctx, inventory.BusUser{Username: p.Username(), Kind: busKindOf(p.Module), Node: p.Node, Module: p.Module})
+			if err != nil {
+				return err
+			}
+			if err := issueWith(ctx, inv, m, p.Node, "", known, busAddress, p.Username(), password); err != nil {
+				return err
+			}
+			modules++
+
+		default:
+			skipped++
+		}
+	}
+	fmt.Printf("minted for %d machine(s) and %d module runtime(s); %d skipped; the bus is at %s\n",
+		machines, modules, skipped, busAddress)
+	fmt.Println("  each machine's membership and each module's credential arrive with the next push of its machine;")
+	fmt.Println("  push the machine running the bus first, so the bus stands with its user list before anything dials it")
+	return nil
+}
+
+// providesBus is whether a manifest provides the mesh's bus.
+func providesBus(m catalogue.Manifest) bool {
+	for _, o := range m.Provides {
+		if o.Name == "mesh-bus" {
+			return true
+		}
+	}
+	return false
+}
+
+// rolloutHand mints a machine its credential for the new bus afresh and prints its membership
+// once, for an operator to carry by hand — the rescue for a machine that cannot be reached over
+// any bus: rotated while it still held the old password, or reachable only by ssh. The plaintext
+// exists on this terminal and then only where it is written; the store keeps the hash, and the
+// sealed copy in the machine's declaration is replaced too, so the next push says the same.
+func rolloutHand(ctx context.Context, node string) error {
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	known, err := broker.FromEnvironment()
+	if err != nil {
+		return fmt.Errorf("the bus's certificate is not known to this process: %w", err)
+	}
+	busAddress, _, err := broker.OnNATS()
+	if err != nil {
+		return err
+	}
+	if busAddress == "" {
+		return errors.New("this control plane is not on the new bus, so there is no membership to hand out")
+	}
+	_, _, bare := broker.CredentialIn(busAddress)
+	if _, after, has := strings.Cut(bare, "://"); has {
+		bare = after
+	}
+	if _, err := inv.NodeByName(ctx, node); err != nil {
+		return err
+	}
+	p := broker.Principal{Kind: broker.KindNode, Node: node}
+	password, err := inv.MintBusPassword(ctx, inventory.BusUser{Username: p.Username(), Kind: inventory.BusNode, Node: node})
+	if err != nil {
+		return err
+	}
+	membership, _ := json.Marshal(map[string]string{
+		"broker": bare, "fingerprint": known.Fingerprint, "password": password, "transport": "nats",
+	})
+	key, err := inv.SealingKeyOf(ctx, node)
+	if err != nil {
+		return err
+	}
+	sealed, err := secrets.Seal(key, membership)
+	if err != nil {
+		return err
+	}
+	if err := inv.PutBusMembership(ctx, node, sealed); err != nil {
+		return err
+	}
+	// The one line of output is the membership itself, so it can be piped to the machine without
+	// being read on the way. Everything else goes to stderr.
+	fmt.Fprintf(os.Stderr, "%s's credential is minted afresh. Write this to %s on it and restart its host; "+
+		"then push the machine running the bus so the user list carries the new hash.\n",
+		node, catalogue.BusMembershipPath)
+	fmt.Println(string(membership))
+	return nil
+}

cmd/mesh-controller/rollout_first_test.go

@@ -0,0 +1,152 @@
+package main
+
+import (
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// novox/hq issue 249, ADR 0218: a plan rolls a module out to one machine first and the rest only
+// once that machine has reported it applied; a module whose policy says together goes everywhere at
+// once, as before.
+func TestAPlanSendsOneMachineFirstAndTheRestAfterItsReport(t *testing.T) {
+	running := []string{"novox", "ace", "g14"}
+	sentAt := time.Date(2026, 10, 5, 12, 0, 0, 0, time.UTC)
+	now := sentAt.Add(5 * time.Minute)
+	bound := 30 * time.Minute
+	after := sentAt.Add(time.Minute)
+	next := func(s inventory.PlanModule, running []string, together bool, reports []inventory.Reported) rolloutStep {
+		return nextRollout(s, running, together, reports, now, bound)
+	}
+
+	// Together: every machine at once.
+	if step := next(inventory.PlanModule{}, running, true, nil); !reflect.DeepEqual(step.send, running) || step.first {
+		t.Fatalf("a together policy did not send every machine at once: %+v", step)
+	}
+
+	// Otherwise the first by name, alone, when none has reported lately.
+	step := next(inventory.PlanModule{}, running, false, nil)
+	if !step.first || !reflect.DeepEqual(step.send, []string{"ace"}) {
+		t.Fatalf("the first send was %+v, wanted ace alone", step)
+	}
+
+	state := inventory.PlanModule{First: []string{"ace"}, FirstAt: &sentAt}
+	report := func(outcome string, current bool) []inventory.Reported {
+		return []inventory.Reported{{Node: "ace", At: &after, Outcome: outcome, Current: current},
+			{Node: "g14", At: &after, Outcome: inventory.OutcomeApplied, Current: true}}
+	}
+
+	// No report yet, or one about an older declaration than it was last sent: wait.
+	for what, reports := range map[string][]inventory.Reported{
+		"no report":            nil,
+		"a report about older": report(inventory.OutcomeApplied, false),
+	} {
+		step := next(state, running, false, reports)
+		if len(step.send) != 0 || step.waiting != "ace" || step.failed != "" {
+			t.Errorf("%s: %+v, wanted to wait for ace", what, step)
+		}
+	}
+
+	// Applied what it was last sent: the rest, and only the rest.
+	step = next(state, running, false, report(inventory.OutcomeApplied, true))
+	if step.first || !reflect.DeepEqual(step.send, []string{"novox", "g14"}) {
+		t.Fatalf("after ace applied it the plan sent %+v, wanted novox and g14", step)
+	}
+
+	// Failed or refused: stop, the rest untouched.
+	for _, outcome := range []string{inventory.OutcomeFailed, inventory.OutcomeRefused} {
+		step := next(state, running, false, report(outcome, true))
+		if len(step.send) != 0 || !strings.Contains(step.failed, "ace "+outcome) ||
+			!reflect.DeepEqual(step.rest, []string{"novox", "g14"}) {
+			t.Errorf("a first machine that %s it: %+v", outcome, step)
+		}
+	}
+
+	// The machine holding the bus went with the first send: the rest wait for it too, and it is
+	// not sent again.
+	both := inventory.PlanModule{First: []string{"novox", "ace"}, FirstAt: &sentAt}
+	half := report(inventory.OutcomeApplied, true)
+	if step := next(both, running, false, half); step.waiting != "novox" {
+		t.Fatalf("the plan did not wait for the bus's machine sent first: %+v", step)
+	}
+	all := append(half, inventory.Reported{Node: "novox", At: &after, Outcome: inventory.OutcomeApplied, Current: true})
+	if step := next(both, running, false, all); !reflect.DeepEqual(step.send, []string{"g14"}) {
+		t.Fatalf("after both applied it the plan sent %+v, wanted g14 alone", step)
+	}
+
+	// One machine, or none: nothing is waited for that cannot come.
+	if step := next(inventory.PlanModule{}, nil, false, nil); len(step.send) != 0 || step.first {
+		t.Fatalf("a module nothing runs was sent: %+v", step)
+	}
+	if step := next(state, []string{"ace"}, false, report(inventory.OutcomeApplied, true)); len(step.send) != 0 || step.waiting != "" {
+		t.Fatalf("a module on one machine waited for more: %+v", step)
+	}
+}
+
+// ADR 0218 §2: a first machine that does not report within the bound stops the rollout there,
+// naming the machine and the bound; the rest are left alone.
+func TestAFirstMachineThatDoesNotReportStopsTheRollout(t *testing.T) {
+	sentAt := time.Date(2026, 10, 5, 12, 0, 0, 0, time.UTC)
+	state := inventory.PlanModule{First: []string{"ace"}, FirstAt: &sentAt}
+	step := nextRollout(state, []string{"ace", "g14"}, false, nil, sentAt.Add(31*time.Minute), 30*time.Minute)
+	if !strings.Contains(step.failed, "ace did not report it applied within 30m") ||
+		!reflect.DeepEqual(step.rest, []string{"g14"}) || len(step.send) != 0 {
+		t.Fatalf("a silent first machine: %+v", step)
+	}
+}
+
+// The first machine is the first by name among those heard from lately: a laptop that is away is
+// not the one the rest wait on. When none has been heard from, the first by name.
+func TestTheFirstMachineIsOneThatHasReportedLately(t *testing.T) {
+	now := time.Date(2026, 10, 5, 12, 0, 0, 0, time.UTC)
+	lately, long := now.Add(-time.Minute), now.Add(-3*time.Hour)
+	reports := []inventory.Reported{
+		{Node: "ace", At: &long}, {Node: "g14", At: &lately}, {Node: "novox", At: &lately},
+	}
+	step := nextRollout(inventory.PlanModule{}, []string{"novox", "ace", "g14"}, false, reports, now, 30*time.Minute)
+	if !step.first || !reflect.DeepEqual(step.send, []string{"g14"}) {
+		t.Fatalf("the first send was %+v, wanted g14, the first heard from lately", step)
+	}
+}
+
+// An announced move of a module an open plan is still rolling out is left to the plan: sending it
+// here as well put the bundle on every machine at once (novox/hq issue 249).
+func TestAnAnnouncedMoveIsLeftToThePlanRollingItOut(t *testing.T) {
+	sent := time.Now()
+	plans := []inventory.Plan{
+		{ID: "plan-done", State: inventory.PlanDone, Modules: map[string]*inventory.PlanModule{"agent": {}}},
+		{ID: "plan-1", State: inventory.PlanRolling, Modules: map[string]*inventory.PlanModule{
+			"agent": {State: "built", First: []string{"ace"}, FirstAt: &sent}, "gitea": {State: "built", SentAt: &sent}}},
+	}
+	if got := rolledOutByAPlan(plans, "agent"); got != "plan-1" {
+		t.Fatalf("a module the plan is rolling out was not left to it: %q", got)
+	}
+	for _, m := range []string{"gitea", "keycloak"} {
+		if got := rolledOutByAPlan(plans, m); got != "" {
+			t.Errorf("%s, which no plan will send, was left to %s", m, got)
+		}
+	}
+}
+
+// A plan that ends without sending what it built says so, with the remedy; a module whose rollout
+// stopped at its first machine is not among them.
+func TestAnEndedPlanSaysWhatItBuiltAndNeverSent(t *testing.T) {
+	at := time.Now()
+	p := inventory.Plan{State: inventory.PlanFailed, Note: "closed by hand at tier 1",
+		Modules: map[string]*inventory.PlanModule{
+			"agent":   {State: "built"},
+			"stopped": {State: "built", First: []string{"ace"}, FirstAt: &at},
+			"sent":    {State: "built", SentAt: &at},
+			"notes":   {State: "built"},
+			"later":   {},
+		}}
+	rollsOut := func(m string) bool { return m != "notes" }
+	sayUnsent(&p, rollsOut)
+	sayUnsent(&p, rollsOut)
+	if p.Note != "closed by hand at tier 1; built and never sent: agent — `push --behind` sends them" {
+		t.Fatalf("the note reads %q", p.Note)
+	}
+}

cmd/mesh-controller/rollout_test.go

@@ -0,0 +1,93 @@
+package main
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// Whether a mesh could move its bus, read from a real store.
+//
+// The readiness reasoning has its own tests; this is about the gathering — that the question is
+// answered from what the mesh actually holds, on a store with machines and modules in it, without
+// writing anything.
+
+func TestReadinessIsGatheredFromWhatTheMeshHolds(t *testing.T) {
+	inv := inventory.ForTest(t)
+	ctx := context.Background()
+
+	// A mesh mid-change: two machines, the bus module on one of them, a module that speaks and a
+	// module that never does.
+	for _, m := range []catalogue.Manifest{
+		{Module: "nats", Version: "1", BusUsers: "/var/lib/nats-module/conf/accounts.conf",
+			Claims: []catalogue.Claim{{Name: "mesh-broker", Scope: catalogue.ScopeMesh}}},
+		{Module: "gitea", Version: "1", Tools: []string{"repo_create"}},
+		{Module: "wallpaper", Version: "1"},
+	} {
+		if err := inv.RegisterModule(ctx, m, inventory.Source{Repository: "/r"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	for _, n := range []string{"anchor", "laptop"} {
+		if _, err := inv.AddNode(ctx, n); err != nil {
+			t.Fatal(err)
+		}
+	}
+	for _, a := range [][2]string{{"anchor", "nats"}, {"anchor", "gitea"}, {"laptop", "wallpaper"}} {
+		if _, err := inv.Assign(ctx, a[0], a[1]); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	state, err := readinessOf(ctx, inv)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if state.Holder != "anchor" {
+		t.Errorf("the machine holding the bus reads as %q", state.Holder)
+	}
+	if len(state.Nodes) != 2 {
+		t.Errorf("machines read as %v", state.Nodes)
+	}
+	// **A module that never speaks is not counted as missing a credential.** A third of the catalogue
+	// never reaches the bus, and listing those would bury the ones that matter.
+	for _, m := range state.Modules {
+		if strings.HasSuffix(m, "/wallpaper") {
+			t.Errorf("a module that never speaks was counted: %v", state.Modules)
+		}
+	}
+	if len(state.Modules) != 2 {
+		t.Errorf("modules that speak read as %v; expected the bus module and the one with a tool",
+			state.Modules)
+	}
+
+	// Nothing has been minted, so it is not ready — and it says so about each machine by name.
+	why := notReadyOf(state)
+	if len(why) == 0 {
+		t.Fatal("a mesh where nothing has a credential was reported ready to move")
+	}
+	said := strings.Join(why, "\n")
+	for _, name := range []string{"anchor", "laptop"} {
+		if !strings.Contains(said, name) {
+			t.Errorf("the refusal does not name %s: %s", name, said)
+		}
+	}
+
+	// Mint for one machine and it drops out of the complaint, which is how somebody works through it.
+	if _, err := inv.MintBusPassword(ctx, inventory.BusUser{
+		Username: "node.laptop", Kind: inventory.BusNode, Node: "laptop",
+	}); err != nil {
+		t.Fatal(err)
+	}
+	state, err = readinessOf(ctx, inv)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !state.Credentialled["laptop"] {
+		t.Error("a machine that was minted a credential still reads as having none")
+	}
+}

cmd/mesh-controller/roster_failure_test.go

@@ -0,0 +1,46 @@
+package main
+
+import (
+	"context"
+	"testing"
+)
+
+// A node's own set failing to compose, and the mesh being unable to answer at all, are different
+// things, and only the first may be passed over when something is gathered across every machine
+// (novox/hq 04-ISSUES/152). These pin that distinction where the gatherers rely on it.
+
+func TestASetThatDoesNotComposeIsMarkedAsTheNodesOwnProblem(t *testing.T) {
+	open := aMesh(t)
+	one, two := rivals()
+	register(t, open, one)
+	register(t, open, two)
+	for _, m := range []string{one.Module, two.Module} {
+		if _, err := open.inventory.Assign(t.Context(), "laptop", m); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	_, _, err := planFor(t.Context(), open, "laptop")
+	if err == nil {
+		t.Fatal("two modules claiming one seat composed anyway")
+	}
+	if !unresolvable(err) {
+		t.Fatalf("a set that cannot compose was not marked as the node's own problem: %v", err)
+	}
+}
+
+func TestAStoreThatCannotBeReadIsNotANodeThatDoesNotCompose(t *testing.T) {
+	open := aMesh(t)
+
+	// Nothing is wrong with anchor. The question simply cannot be asked.
+	stopped, cancel := context.WithCancel(t.Context())
+	cancel()
+
+	_, _, err := planFor(stopped, open, "anchor")
+	if err == nil {
+		t.Fatal("a plan composed against a store that could not be read")
+	}
+	if unresolvable(err) {
+		t.Fatalf("a question the mesh could not answer was read as a node that runs nothing: %v", err)
+	}
+}

cmd/mesh-controller/rotate.go

@@ -0,0 +1,125 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"sort"
+)
+
+// rotateCommand replaces a credential and moves both ends together.
+//
+// **This is the invariant novox/hq ADR 0001 records as unowned, and it was measurably false.** On
+// 2026-08-22 `provision_ensure` — documented as never rotating an existing secret — minted a new
+// password on every adoption and updated only the provider's row. Consumers on three nodes held
+// dead credentials for two days; two rows for one provision were written 216 ms apart, so at most
+// one could match the live role. Nothing enumerated who held the old one, and nothing said so.
+//
+// Three things make that impossible here, and all three are deliberate:
+//
+// **The holders are a set the mesh can name.** Each pair has its own credential, so rotating one
+// consumer's password touches one role and leaves every other consumer alone — and the list of who
+// is affected is a query rather than an assumption.
+//
+// **Both ends are pushed by this command, not by a later one.** A rotation that changed the record
+// and left the sending to whoever remembered is the fault above, exactly.
+//
+// **It is all-or-nothing.** If any affected machine cannot be resolved, nothing is sent and the old
+// credential keeps working — which is a mesh that has not rotated, and is far better than one that
+// has half-rotated.
+func rotateCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("rotate", flag.ContinueOnError)
+	// One consumer rather than all of them. Ordinary: a credential is suspected on one machine,
+	// and rotating the other nine would be a great deal of disruption for one suspicion.
+	only := set.String("consumer", "", "only this machine's credential, rather than every holder's")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) != 1 {
+		return errors.New("rotate <provision> [--consumer <machine>]")
+	}
+	provision := positionals[0]
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	holders, err := inv.HoldersOf(ctx, provision, *only)
+	if err != nil {
+		return err
+	}
+	if len(holders) == 0 {
+		// Said, not silent. "Nobody holds this" and "this did not run" must never look the same —
+		// and a rotation somebody believes happened is worse than one they know did not.
+		if *only != "" {
+			return fmt.Errorf(
+				"%s holds no credential for %q, so there is nothing to rotate. `plan %s` says "+
+					"what it does hold", *only, provision, *only)
+		}
+		return fmt.Errorf(
+			"nothing in this mesh holds a credential for %q, so there is nothing to rotate",
+			provision)
+	}
+
+	// Every machine at both ends, named before anything changes. A person about to rotate a
+	// production credential is entitled to know the blast radius before it is the past tense.
+	affected := map[string]bool{}
+	for _, h := range holders {
+		affected[h.Consumer] = true
+		affected[h.Provider] = true
+	}
+	machines := make([]string, 0, len(affected))
+	for name := range affected {
+		machines = append(machines, name)
+	}
+	sort.Strings(machines)
+
+	fmt.Printf("rotating %s for %d holder(s):\n", provision, len(holders))
+	for _, h := range holders {
+		// The module, because a machine may hold several credentials for one provision and
+		// rotating "anchor's database password" now means rotating three of them.
+		fmt.Printf("  %s on %s, from %s%s\n", h.ConsumerModule, h.Consumer, h.Provider, asLocal(h.Local))
+	}
+
+	for _, h := range holders {
+		if err := inv.RotateSecret(ctx, h.Provision, h.Consumer, h.ConsumerModule, h.Provider, h.Local); err != nil {
+			// Partly rotated, and said so plainly. What is gone is remade on the next push, so
+			// the remedy is to run this again rather than to repair anything — but a machine
+			// whose secret was discarded and not resent is holding a credential the provider is
+			// about to stop honouring, and that is worth knowing now.
+			return fmt.Errorf(
+				"rotating %s for %s on %s from %s: %w\n\nSome credentials were discarded and "+
+					"not yet sent. Run this again once the cause is fixed",
+				h.Provision, h.ConsumerModule, h.Consumer, h.Provider, err)
+		}
+	}
+
+	// **Both ends, in one send.** There is a window either way — a role's password changes on the
+	// provider and the file changes on the consumer, and they cannot be simultaneous — so the
+	// honest thing is to make it as short as the broker allows and to never leave it open across
+	// a command boundary, where it depends on somebody's memory.
+	fmt.Printf("\nsending to both ends:\n")
+	if err := sendTo(ctx, open, machines); err != nil {
+		return fmt.Errorf(
+			"%w\n\nThe old credentials are gone from the mesh and the new ones have not arrived. "+
+				"Nothing on those machines has changed yet, so what is running keeps working "+
+				"until the provider next applies. Fix the cause and run `push --behind`", err)
+	}
+
+	fmt.Printf("\n%d machine(s) told. Until both ends have applied, a consumer whose password "+
+		"changed cannot authenticate — `status` says who is still behind\n", len(machines))
+	return nil
+}
+
+// asLocal names the credential inside the consumer where it holds several (ADR 0094).
+func asLocal(local string) string {
+	if local == "" {
+		return ""
+	}
+	return " (as " + local + ")"
+}

cmd/mesh-controller/seat_dependencies_test.go

@@ -0,0 +1,149 @@
+package main
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// Defends novox/hq ADR 0207 at the controller's acts: `assign` refuses a module whose resources a
+// seat nothing on the node holds applies, `unassign` refuses taking the last holder from under its
+// dependents, and `status` reports what composition does not yet refuse.
+
+func serviceManagerHolder() catalogue.Manifest {
+	return catalogue.Manifest{Module: "systemd", Version: "1",
+		Claims: []catalogue.Claim{{Name: catalogue.ServiceManagerSeat, Scope: catalogue.ScopeNode,
+			Serves: []string{"units", "status", "start", "stop", "restart", "enable", "disable", "journal"}}},
+		Resources: []map[string]any{{"id": "systemd", "type": "package", "package": "systemd"}}}
+}
+
+func packageManagerHolder() catalogue.Manifest {
+	return catalogue.Manifest{Module: "pacman", Version: "1",
+		Claims:    []catalogue.Claim{{Name: catalogue.PackageManagerSeat, Scope: catalogue.ScopeNode}},
+		Resources: []map[string]any{{"id": "refresh", "type": "service", "unit": "pacman-refresh.timer"}}}
+}
+
+func aDaemon() catalogue.Manifest {
+	return catalogue.Manifest{Module: "sshd", Version: "1",
+		Resources: []map[string]any{{"id": "sshd", "type": "service", "unit": "sshd.service"}}}
+}
+
+func TestAnAssignmentWithoutItsHolderIsRefusedAndNotKept(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, serviceManagerHolder())
+	register(t, open, packageManagerHolder())
+	register(t, open, aDaemon())
+
+	_, err := assign(ctx, open, "laptop", "sshd")
+	if err == nil {
+		t.Fatal("sshd went onto a machine nothing holds the service manager of")
+	}
+	for _, want := range []string{catalogue.ServiceManagerSeat, "systemd", "ADR 0207"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Errorf("the refusal does not say %q:\n%v", want, err)
+		}
+	}
+	assigned, err := open.inventory.Assigned(ctx, "laptop")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if contains(assigned, "sshd") {
+		t.Fatalf("a refused assignment was kept: %v", assigned)
+	}
+
+	// The holders depend on each other, so neither goes on alone — and both go on in one act.
+	if _, err := assign(ctx, open, "laptop", "systemd"); err == nil {
+		t.Fatal("systemd went on alone though its package needs a package manager")
+	}
+	if said, err := assign(ctx, open, "laptop", "systemd", "pacman"); err != nil {
+		t.Fatalf("the two holders assigned together were refused: %v\n%s", err, said)
+	}
+	if said, err := assign(ctx, open, "laptop", "sshd"); err != nil {
+		t.Fatalf("sshd beside its holder was refused: %v\n%s", err, said)
+	}
+}
+
+func TestTheControllerSeatsAssignTakesSeveralModulesAsOneAct(t *testing.T) {
+	argv, err := argvFor("assign", map[string]any{"node": "laptop", "module": "systemd, pacman"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Join(argv, " ") != "assign laptop systemd pacman" {
+		t.Errorf("the seat's assign became %v", argv)
+	}
+}
+
+func TestUnassigningTheLastHolderUnderItsDependentsIsRefused(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, serviceManagerHolder())
+	register(t, open, packageManagerHolder())
+	register(t, open, aDaemon())
+	if _, err := assign(ctx, open, "laptop", "systemd", "pacman", "sshd"); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := unassign(ctx, open, "laptop", "systemd")
+	if err == nil {
+		t.Fatal("the service manager came off a machine still running services")
+	}
+	for _, want := range []string{catalogue.ServiceManagerSeat, "sshd", "pacman"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Errorf("the refusal does not name %q:\n%v", want, err)
+		}
+	}
+	assigned, _ := open.inventory.Assigned(ctx, "laptop")
+	if !contains(assigned, "systemd") {
+		t.Fatalf("a refused unassignment took the module off anyway: %v", assigned)
+	}
+	if _, err := unassign(ctx, open, "laptop", "sshd"); err != nil {
+		t.Fatalf("a dependent could not come off: %v", err)
+	}
+}
+
+func TestStatusReportsAnUnheldDependencyWithoutRefusingTheMachine(t *testing.T) {
+	// The mesh as it ran before the switch (novox/hq ADR 0207 §4).
+	defer catalogue.EnforcingSeatDependencies(false)()
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, serviceManagerHolder())
+	register(t, open, aDaemon())
+	// Assigned straight into the store: a machine whose modules predate the rule, which is every
+	// machine on the day it ships.
+	if _, err := open.inventory.Assign(ctx, "laptop", "sshd"); err != nil {
+		t.Fatal(err)
+	}
+
+	asked, err := theThreeQuestions(ctx, open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, refused := asked.refused["laptop"]; refused {
+		t.Fatalf("an unmet dependency refused the machine before the switch: %s", asked.refused["laptop"])
+	}
+	if asked.well() {
+		t.Error("a mesh with an unheld dependency reads as all well")
+	}
+	got := printed(t, func() error { return printStatus(asked) })
+	for _, want := range []string{"unheld", "laptop", "sshd", catalogue.ServiceManagerSeat, "systemd"} {
+		if !strings.Contains(got, want) {
+			t.Errorf("status does not say %q:\n%s", want, got)
+		}
+	}
+	body, err := statusAsJSON(asked)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var doc struct {
+		Unheld []catalogue.Unheld `json:"unheld"`
+	}
+	if err := json.Unmarshal(body, &doc); err != nil {
+		t.Fatal(err)
+	}
+	if len(doc.Unheld) != 1 || doc.Unheld[0].Module != "sshd" || doc.Unheld[0].Seat != catalogue.ServiceManagerSeat {
+		t.Errorf("the document's unheld is %+v", doc.Unheld)
+	}
+}

cmd/mesh-controller/seats.go

@@ -0,0 +1,293 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"flag"
+	"fmt"
+	"os"
+	"slices"
+	"sort"
+	"strings"
+	"text/tabwriter"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// What this mesh can have one of, and who fills each (novox/hq ADR 0110).
+//
+// **Derived every time, never stored.** A seat is held by a module assignment, so the answer is
+// computed from assignments by the same resolution that decides what every machine runs. A table
+// of holders kept beside the assignments would be a second copy of one fact, and the first thing
+// to be wrong about it.
+
+// seatHolder is one assignment holding a seat.
+type seatHolder struct {
+	Node   string `json:"node"`
+	Module string `json:"module"`
+}
+
+// seatRow is one seat and who holds it. Unheld is an answer — "this mesh has no X" — not a fault.
+type seatRow struct {
+	Seat     string `json:"seat"`
+	Scope    string `json:"scope"`
+	Delivers string `json:"delivers,omitempty"`
+	Decision string `json:"decision"`
+	// Replicated says the seat may be held on several machines at once (novox/hq ADR 0223).
+	Replicated bool         `json:"replicated,omitempty"`
+	Holders    []seatHolder `json:"holders"`
+}
+
+// seatsHeld is every seat the mesh defines with its holders, and every claim held that names no
+// seat in the set.
+//
+// **The second list is not empty by construction.** Manifests are held to the set when they are
+// registered, and a mesh can hold one registered before the set closed. Leaving its claim out of the
+// overview would make the one thing the overview is for — what does this mesh have — quietly
+// incomplete.
+func seatsHeld(seats []catalogue.Seat, held []catalogue.Held) ([]seatRow, []catalogue.Held) {
+	rows := make([]seatRow, 0, len(seats))
+	for _, s := range seats {
+		row := seatRow{Seat: s.Name, Scope: s.Scope, Delivers: s.Delivers, Decision: s.Decision,
+			Replicated: s.Replicated, Holders: []seatHolder{}}
+		seen := map[seatHolder]bool{}
+		for _, h := range held {
+			// Resolve the held claim to a seat rather than comparing names, so a record naming a
+			// seat's former name groups under it after a rename (novox/hq ADR 0122).
+			hs, ok := catalogue.SeatNamed(h.Claim)
+			if !ok || hs.Name != s.Name || h.Scope != s.Scope {
+				continue
+			}
+			holder := seatHolder{Node: h.Node, Module: h.Module}
+			if !seen[holder] {
+				seen[holder] = true
+				row.Holders = append(row.Holders, holder)
+			}
+		}
+		sort.Slice(row.Holders, func(i, j int) bool {
+			if row.Holders[i].Node != row.Holders[j].Node {
+				return row.Holders[i].Node < row.Holders[j].Node
+			}
+			return row.Holders[i].Module < row.Holders[j].Module
+		})
+		rows = append(rows, row)
+	}
+	var outside []catalogue.Held
+	for _, h := range held {
+		// Outside the set only if it resolves to no seat at all — a former name still resolves.
+		if _, ok := catalogue.SeatNamed(h.Claim); !ok {
+			outside = append(outside, h)
+		}
+	}
+	sort.Slice(outside, func(i, j int) bool {
+		if outside[i].Claim != outside[j].Claim {
+			return outside[i].Claim < outside[j].Claim
+		}
+		return outside[i].Node < outside[j].Node
+	})
+	return rows, outside
+}
+
+// seatCommand changes the set — the whole point of it being data (novox/hq ADR 0122) — and, since
+// ADR 0131, changes who holds a seat.
+func seatCommand(ctx context.Context, args []string) error {
+	if len(args) == 3 && args[0] == "rename" {
+		from, to := args[1], args[2]
+		open, err := openStores(ctx)
+		if err != nil {
+			return err
+		}
+		defer open.Close()
+		if err := open.inventory.RenameSeat(ctx, from, to); err != nil {
+			return err
+		}
+		fmt.Printf("%s is now %s — its former name still resolves, so nothing is rebuilt, "+
+			"re-registered or frozen (novox/hq ADR 0122)\n", from, to)
+		return nil
+	}
+	if len(args) == 3 && args[1] == "--to" {
+		return handOver(ctx, args[0], args[2], false)
+	}
+	if len(args) == 3 && args[1] == "--add" {
+		return handOver(ctx, args[0], args[2], true)
+	}
+	return fmt.Errorf("seat rename <from> <to>  |  seat <name> --to <node>/<module>  |  " +
+		"seat <name> --add <node>/<module>")
+}
+
+// handOver makes one assignment the holder of a seat, as one act, so the seat is never without a
+// holder in between (novox/hq ADR 0131, design 28 task 5.3). The control plane finds its own bus
+// through one of these seats; the day it was left empty mid-change is why this exists.
+//
+// Everything that could make the new holder wrong is refused here, before the row is written: the
+// seat must exist, the assignment must exist, and the module must be able to hold the seat —
+// claim it at its scope and provide what it delivers, judged against the store's row. What is
+// **not** checked is whether the module is running yet: that is what `push` confirms afterwards,
+// and refusing to record a handover to a module the node has not started would make the handover
+// impossible to do before the switch instead of as the switch.
+//
+// **Or adds one holder beside the others, for a replicated seat** (novox/hq ADR 0223): `--add`
+// records the named assignment as a further holder and leaves every holder on record as it is. A
+// seat held once refuses it, naming `--to`; `--to` on a replicated seat replaces every holder with
+// the one named, as it always did.
+func handOver(ctx context.Context, seatName, to string, adding bool) error {
+	nodeName, module, ok := strings.Cut(to, "/")
+	if !ok || nodeName == "" || module == "" {
+		return fmt.Errorf("the new holder is named <node>/<module>, not %q", to)
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	seat, known := catalogue.SeatNamed(seatName)
+	if !known {
+		return fmt.Errorf("%q is not a seat this mesh defines — `seats` lists them", seatName)
+	}
+	if adding && !seat.Replicated {
+		return fmt.Errorf("%s is held once per %s, so a second holder cannot be added beside the first — "+
+			"`seat %s --to %s` hands it over", seat.Name, seat.Scope, seat.Name, to)
+	}
+	assigned, err := inv.Assigned(ctx, nodeName)
+	if err != nil {
+		return err
+	}
+	if !slices.Contains(assigned, module) {
+		return fmt.Errorf("%s is not assigned to %s, so it cannot hold anything there — "+
+			"`assign %s %s` first", module, nodeName, nodeName, module)
+	}
+	entries, err := inv.Catalogued(ctx)
+	if err != nil {
+		return err
+	}
+	var m *catalogue.Manifest
+	for i := range entries {
+		if entries[i].Manifest.Module == module {
+			m = &entries[i].Manifest
+		}
+	}
+	if m == nil {
+		return fmt.Errorf("%s is assigned but not in the catalogue, which should not happen", module)
+	}
+	var was string
+	var held []string
+	holdings, err := inv.Holdings(ctx)
+	if err != nil {
+		return err
+	}
+	for _, h := range holdings {
+		if hs, ok := catalogue.SeatNamed(h.Claim); ok && hs.Name == seat.Name {
+			was = h.Node
+			held = append(held, h.Node)
+		}
+	}
+
+	// **Recording who already holds the seat is not making a new holder, and is not judged like
+	// one.** On a mesh that predates the record, the first handover has to begin by writing down
+	// the standing holder — otherwise the next holder cannot be assigned beside it, because two
+	// eligible claimants with nothing on record are refused. That standing holder may no longer
+	// satisfy what the seat delivers (the row moved under it, on purpose, as ADR 0131's first step),
+	// and it holds regardless: derivation never read that column. So when nothing is on record and
+	// the named assignment is the one holding by derivation, only the claim itself is checked here.
+	// Every *change* of holder is judged in full.
+	claimsIt := false
+	for _, c := range m.Claims {
+		if cs, ok := catalogue.SeatNamed(c.Name); ok && cs.Name == seat.Name && c.At() == seat.Scope {
+			claimsIt = true
+		}
+	}
+	if was == "" && claimsIt {
+		fmt.Printf("nothing was on record for %s; recording %s on %s as its standing holder\n",
+			seat.Name, module, nodeName)
+	} else if err := catalogue.CanHold(*m, seat); err != nil {
+		return fmt.Errorf("%s cannot hold %s: %w", module, seat.Name, err)
+	}
+	if adding {
+		if err := inv.AddSeatHolder(ctx, seat.Name, seat.Scope, nodeName, module); err != nil {
+			return err
+		}
+		fmt.Printf("%s is held by %s on %s, beside what was on record: %s\n", seat.Name, module, nodeName,
+			strings.Join(held, ", "))
+		fmt.Printf("  `push --behind` re-declares every machine that reads the seat's holders\n")
+		return nil
+	}
+	if err := inv.HoldSeat(ctx, seat.Name, seat.Scope, nodeName, module); err != nil {
+		return err
+	}
+	fmt.Printf("%s is held by %s on %s\n", seat.Name, module, nodeName)
+	if was != "" && was != nodeName {
+		fmt.Printf("  `push %s` and `push %s` send both machines what changed\n", was, nodeName)
+	} else {
+		fmt.Printf("  `push %s` sends the machine what changed; every other machine that reads the "+
+			"seat is re-declared by `push --behind`\n", nodeName)
+	}
+	return nil
+}
+
+func seatsCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("seats", flag.ContinueOnError)
+	asJSON := set.Bool("json", false, "the same, as JSON")
+	if err := set.Parse(args); err != nil {
+		return err
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return err
+	}
+	// Every node, none excluded: the same view of what each machine holds that planning uses.
+	world, err := theRestOfTheMesh(ctx, inv, shelf, "")
+	if err != nil {
+		return err
+	}
+	rows, outside := seatsHeld(catalogue.Seats(), world.Held)
+
+	if *asJSON {
+		out := struct {
+			Seats   []seatRow        `json:"seats"`
+			Outside []catalogue.Held `json:"outside,omitempty"`
+		}{rows, outside}
+		body, err := json.MarshalIndent(out, "", "  ")
+		if err != nil {
+			return err
+		}
+		fmt.Println(string(body))
+		return nil
+	}
+
+	w := tabwriter.NewWriter(os.Stdout, 0, 0, 2, ' ', 0)
+	fmt.Fprintln(w, "SEAT\tSCOPE\tDELIVERS\tHELD BY")
+	for _, r := range rows {
+		delivers := r.Delivers
+		if delivers == "" {
+			delivers = "—"
+		}
+		holders := "unheld"
+		if len(r.Holders) > 0 {
+			parts := make([]string, 0, len(r.Holders))
+			for _, h := range r.Holders {
+				parts = append(parts, h.Module+" on "+h.Node)
+			}
+			holders = strings.Join(parts, ", ")
+		}
+		fmt.Fprintf(w, "%s\t%s\t%s\t%s\n", r.Seat, r.Scope, delivers, holders)
+	}
+	if err := w.Flush(); err != nil {
+		return err
+	}
+	if len(outside) > 0 {
+		fmt.Println("\nheld, and not a seat this mesh defines (registered before the set closed — novox/hq ADR 0110):")
+		for _, h := range outside {
+			fmt.Printf("  %s  %s on %s\n", h.Claim, h.Module, h.Node)
+		}
+	}
+	return nil
+}

cmd/mesh-controller/seats_test.go

@@ -0,0 +1,64 @@
+package main
+
+import (
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// The overview of what a mesh has (novox/hq ADR 0110).
+
+func TestEverySeatIsListedIncludingTheOnesNobodyHolds(t *testing.T) {
+	// An unheld seat is an answer — "this mesh has no forge" — so it is listed rather than omitted.
+	rows, _ := seatsHeld(catalogue.Seats(), []catalogue.Held{
+		{Claim: "mesh-store", Scope: catalogue.ScopeMesh, Node: "anchor", Module: "postgres"},
+	})
+	if len(rows) != len(catalogue.Seats()) {
+		t.Fatalf("%d seats listed of %d", len(rows), len(catalogue.Seats()))
+	}
+	for _, r := range rows {
+		switch r.Seat {
+		case "mesh-store":
+			if len(r.Holders) != 1 || r.Holders[0].Module != "postgres" || r.Holders[0].Node != "anchor" {
+				t.Errorf("mesh-store is held by %+v", r.Holders)
+			}
+			if r.Delivers != "postgres-database" {
+				t.Errorf("mesh-store does not say what it delivers: %q", r.Delivers)
+			}
+		case "git":
+			if len(r.Holders) != 0 {
+				t.Errorf("git is held by %+v in a mesh with no forge", r.Holders)
+			}
+		}
+	}
+}
+
+func TestANodeSeatListsEveryMachineHoldingIt(t *testing.T) {
+	rows, _ := seatsHeld(catalogue.Seats(), []catalogue.Held{
+		{Claim: "node-packet-filter", Scope: catalogue.ScopeNode, Node: "node2", Module: "nftables"},
+		{Claim: "node-packet-filter", Scope: catalogue.ScopeNode, Node: "anchor", Module: "nftables"},
+		// Resolved twice, reported once: a machine is one holder however many passes saw it.
+		{Claim: "node-packet-filter", Scope: catalogue.ScopeNode, Node: "anchor", Module: "nftables"},
+	})
+	for _, r := range rows {
+		if r.Seat != "node-packet-filter" {
+			continue
+		}
+		if len(r.Holders) != 2 || r.Holders[0].Node != "anchor" || r.Holders[1].Node != "node2" {
+			t.Fatalf("the packet filter is held by %+v", r.Holders)
+		}
+		return
+	}
+	t.Fatal("the packet filter is not listed")
+}
+
+func TestAClaimOutsideTheSetIsShownNotHidden(t *testing.T) {
+	// A manifest registered before the set closed can still hold one. Leaving it out would make
+	// the overview quietly incomplete, which is the one thing it may not be.
+	_, outside := seatsHeld(catalogue.Seats(), []catalogue.Held{
+		{Claim: "the-controller", Scope: catalogue.ScopeMesh, Node: "anchor", Module: "mesh-controller"},
+	})
+	if len(outside) != 1 || outside[0].Claim != "the-controller" {
+		t.Fatalf("a claim outside the set was not shown: %+v", outside)
+	}
+}

cmd/mesh-controller/seatverbs.go

@@ -0,0 +1,718 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"github.com/nats-io/nats.go/micro"
+	"os"
+	"os/exec"
+	"sort"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// The mesh's own verbs, served as the mesh-controller seat's tools (novox/hq ADR 0154, design 33).
+//
+// **Each tool runs the command it names, in this same binary, and answers what it printed.** That is
+// ADR 0035 taken literally: the logic lives once, in the command, and a surface is an adapter with no
+// decisions in it. Running a fresh process rather than calling the function keeps two things true
+// that calling it would not — every command opens and closes its own stores the way it does from a
+// shell, and nothing a command prints to the process's standard output can leak into another call's
+// answer. It also means a refusal is the same refusal in the same words, because it is the same
+// output.
+
+// verbAnswer is what a verb answers: what the command printed, whether it succeeded, and — where the
+// command speaks JSON — the same as data.
+type verbAnswer struct {
+	Output string `json:"output"`
+	OK     bool   `json:"ok"`
+	Answer any    `json:"answer,omitempty"`
+}
+
+// argvFor is the command line a verb and its arguments become. Only the verbs the seat declares, and
+// only the arguments each declares: a caller cannot reach a flag the schema did not name.
+//
+// **Nothing a caller sends is passed over** (novox/hq issue 244). An argument the verb does not
+// declare is refused, naming it; a switch that is not "true" or "false" is refused; and an argument
+// the verb declares but did not use for the command line it composed — given beside another that
+// wins, or half of a shape — is refused too. On 2026-10-05 a push naming one machine reached the
+// verb without the machine and ran as a push of every machine behind; a verb that answers "I did
+// not take that" would have stopped it before anything was sent.
+func argvFor(verb string, args map[string]any) ([]string, error) {
+	a, err := readArguments(verb, args)
+	if err != nil {
+		return nil, err
+	}
+	argv, err := a.commandLine()
+	if len(a.misread) > 0 {
+		// The table and the command line disagree: the verb reads an argument no caller can see
+		// in its schema, so no caller could ever pass it.
+		return nil, fmt.Errorf("%s reads %s, which its schema does not declare — this build's verb "+
+			"table and its command lines disagree", verb, quoteAll(a.misread))
+	}
+	if err != nil {
+		return nil, err
+	}
+	if unused := a.unused(); len(unused) > 0 {
+		return nil, fmt.Errorf("%s did not use %s together with %s, and an argument a verb would pass over "+
+			"is refused: nothing was done", verb, quoteAll(unused), quoteAll(a.usedGiven()))
+	}
+	return argv, nil
+}
+
+// verbArguments are one call's arguments, checked against the verb's schema, and which of them the
+// command line was composed from.
+type verbArguments struct {
+	verb     string
+	given    map[string]string
+	used     map[string]bool
+	declared map[string]bool
+	misread  []string // arguments the command line read that the schema does not declare: a bug here
+}
+
+// controllerVerb is this binary's own definition of a verb: what it runs is what it declares, so the
+// arguments are checked against the table compiled beside argvFor, not a row a newer or older build
+// wrote.
+func controllerVerb(name string) (catalogue.Verb, bool) {
+	for _, v := range catalogue.ControllerVerbs {
+		if v.Name == name {
+			return v, true
+		}
+	}
+	return catalogue.Verb{}, false
+}
+
+// declaredArguments are a schema's properties, and which of them are switches.
+func declaredArguments(v catalogue.Verb) (names []string, switches map[string]bool) {
+	switches = map[string]bool{}
+	props, _ := v.Input["properties"].(map[string]any)
+	for name, p := range props {
+		names = append(names, name)
+		desc, _ := p.(map[string]any)
+		switch enum := desc["enum"].(type) {
+		case []string:
+			switches[name] = len(enum) == 2 && enum[0] == "true" && enum[1] == "false"
+		case []any:
+			switches[name] = len(enum) == 2 && enum[0] == "true" && enum[1] == "false"
+		}
+	}
+	sort.Strings(names)
+	return names, switches
+}
+
+// readArguments refuses what the verb does not take, before anything is composed.
+func readArguments(verb string, args map[string]any) (*verbArguments, error) {
+	v, known := controllerVerb(verb)
+	if !known {
+		return nil, fmt.Errorf("%q is not a verb the %s seat serves", verb, catalogue.ControllerSeatName)
+	}
+	names, switches := declaredArguments(v)
+	declared := map[string]bool{}
+	for _, n := range names {
+		declared[n] = true
+	}
+	takes := "none"
+	if len(names) > 0 {
+		takes = quoteAll(names)
+	}
+	a := &verbArguments{verb: verb, given: map[string]string{}, used: map[string]bool{}, declared: declared}
+	keys := make([]string, 0, len(args))
+	for k := range args {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+	for _, k := range keys {
+		if !declared[k] {
+			return nil, fmt.Errorf("%s takes no argument %q — it takes %s; nothing was done", verb, k, takes)
+		}
+		var value string
+		switch x := args[k].(type) {
+		case nil:
+			continue
+		case string:
+			value = strings.TrimSpace(x)
+		case bool:
+			if !switches[k] {
+				return nil, fmt.Errorf("%s: %q is text, not true or false", verb, k)
+			}
+			value = fmt.Sprint(x)
+		default:
+			return nil, fmt.Errorf("%s: %q is text, and was given %T", verb, k, x)
+		}
+		if switches[k] {
+			switch value {
+			case "true":
+			case "false", "":
+				continue // said and off: the same as not given, and nothing passed over
+			default:
+				return nil, fmt.Errorf("%s: %q is \"true\" or \"false\", not %q", verb, k, value)
+			}
+		}
+		if value != "" {
+			a.given[k] = value
+		}
+	}
+	return a, nil
+}
+
+// str is one argument's value, marked as used.
+func (a *verbArguments) str(key string) string {
+	if !a.declared[key] {
+		a.misread = append(a.misread, key)
+	}
+	a.used[key] = true
+	return a.given[key]
+}
+
+// on is a switch, marked as used.
+func (a *verbArguments) on(key string) bool { return a.str(key) == "true" }
+
+// need refuses a call missing a required argument, in the verb's own words.
+func (a *verbArguments) need(keys ...string) error {
+	for _, k := range keys {
+		if a.str(k) == "" {
+			return fmt.Errorf("%s needs %q", a.verb, k)
+		}
+	}
+	return nil
+}
+
+// unused are the arguments given that the command line was not composed from.
+func (a *verbArguments) unused() []string {
+	var out []string
+	for k := range a.given {
+		if !a.used[k] {
+			out = append(out, k)
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+func (a *verbArguments) usedGiven() []string {
+	var out []string
+	for k := range a.given {
+		if a.used[k] {
+			out = append(out, k)
+		}
+	}
+	sort.Strings(out)
+	if len(out) == 0 {
+		return []string{"nothing"}
+	}
+	return out
+}
+
+func quoteAll(xs []string) string {
+	q := make([]string, len(xs))
+	for i, x := range xs {
+		if x == "nothing" {
+			q[i] = x
+			continue
+		}
+		q[i] = fmt.Sprintf("%q", x)
+	}
+	return strings.Join(q, ", ")
+}
+
+// commandLine composes the command. Every argument it reads is one it uses: a branch that reads an
+// argument and then drops it would pass it over, which is what the check after it exists to refuse.
+func (a *verbArguments) commandLine() ([]string, error) {
+	verb, str, on, need := a.verb, a.str, a.on, a.need
+	switch verb {
+	case "command":
+		// The generic verb: the command line as given, split as a shell would split it, with
+		// nothing added — the named verbs add flags a caller cannot reach; this one is the whole
+		// binary and says so in its description (novox/hq ADR 0154, 0175).
+		if err := need("command"); err != nil {
+			return nil, err
+		}
+		argv, err := splitCommandLine(str("command"))
+		if err != nil {
+			return nil, err
+		}
+		if len(argv) == 0 {
+			return nil, errors.New("command names no command")
+		}
+		return argv, nil
+	case "tools":
+		return nil, errors.New("tools is answered from the records, not by a command")
+	case "status":
+		return []string{"status", "--json"}, nil
+	case "nodes":
+		return []string{"node", "list", "--json"}, nil
+	case "node":
+		if err := need("node"); err != nil {
+			return nil, err
+		}
+		return []string{"node", "show", str("node")}, nil
+	case "modules":
+		return []string{"module", "list", "--json"}, nil
+	case "seats":
+		return []string{"seats", "--json"}, nil
+	case "builds":
+		if id := str("log"); id != "" {
+			return []string{"builds", "--log", id}, nil
+		}
+		argv := []string{"builds"}
+		if n := str("limit"); n != "" {
+			argv = append(argv, "-n", n)
+		}
+		if m := str("module"); m != "" {
+			argv = append(argv, m)
+		}
+		return argv, nil
+	case "plans":
+		if r := str("repository"); r != "" {
+			argv := []string{"plans", "--what-if", r}
+			if p := str("paths"); p != "" {
+				argv = append(argv, "--paths", p)
+			}
+			if m := str("modules"); m != "" {
+				argv = append(argv, "--modules", m)
+			}
+			return argv, nil
+		}
+		for _, act := range []string{"stop", "close", "retry"} {
+			if id := str(act); id != "" {
+				return []string{"plans", act, id}, nil
+			}
+		}
+		if id := str("id"); id != "" {
+			return []string{"plans", id}, nil
+		}
+		argv := []string{"plans"}
+		if n := str("limit"); n != "" {
+			argv = append(argv, "-n", n)
+		}
+		return argv, nil
+	// The build queue (novox/hq ADR 0219).
+	case "queue":
+		return []string{"queue"}, nil
+	case "cancel", "kill":
+		if err := need("id"); err != nil {
+			return nil, err
+		}
+		return []string{verb, str("id")}, nil
+	case "clear":
+		if on("dead") {
+			return []string{"clear", "--dead"}, nil
+		}
+		return []string{"clear"}, nil
+	case "rebuild":
+		if err := need("what"); err != nil {
+			return nil, err
+		}
+		return []string{"rebuild", str("what")}, nil
+	case "replay":
+		if err := need("id"); err != nil {
+			return nil, err
+		}
+		argv := []string{"replay", str("id")}
+		if on("register") {
+			argv = append(argv, "--register")
+		}
+		if on("older") {
+			argv = append(argv, "--older")
+		}
+		return argv, nil
+	case "pause", "resume":
+		if n := str("node"); n != "" {
+			return []string{verb, n}, nil
+		}
+		return []string{verb}, nil
+	case "plan":
+		if err := need("node"); err != nil {
+			return nil, err
+		}
+		if on("files") {
+			return []string{"plan", str("node"), "--files"}, nil
+		}
+		return []string{"plan", str("node"), "--json"}, nil
+	case "assign", "unassign":
+		if err := need("node", "module"); err != nil {
+			return nil, err
+		}
+		// Several modules comma-separated, judged as one act (novox/hq ADR 0207): the holders of
+		// the seats that apply resources depend on each other and go on together.
+		return append([]string{verb, str("node")}, splitModules(str("module"))...), nil
+	case "pin":
+		if err := need("node", "provision", "from", "module"); err != nil {
+			return nil, err
+		}
+		return []string{"pin", str("node"), str("provision"), str("from"), str("module")}, nil
+	case "unpin":
+		if err := need("node", "provision"); err != nil {
+			return nil, err
+		}
+		return []string{"unpin", str("node"), str("provision")}, nil
+	case "push":
+		// Sent and not waited for: the asker reads `status` for what the machine did, which is
+		// what a person at a shell does too. A tool call that blocked for a push's whole apply would
+		// time out on every machine that takes a minute, and say nothing about the ones that did not.
+		if n := str("node"); n != "" {
+			// behind is not read here: given with a machine, it is refused as passed over — naming
+			// a machine and asking for every machine behind are two requests, and guessing one
+			// would push a machine nobody named, or not push one somebody did.
+			return []string{"push", n, "--wait", "0"}, nil
+		}
+		// No machine: the whole mesh, whether or not behind said so. The command's answer says it
+		// first, so a caller who meant one machine reads that it was not one.
+		on("behind")
+		return []string{"push", "--behind", "--wait", "0"}, nil
+	case "rotate":
+		if p := str("provision"); p != "" {
+			argv := []string{"rotate", p}
+			if c := str("consumer"); c != "" {
+				argv = append(argv, "--consumer", c)
+			}
+			return argv, nil
+		}
+		_, node := a.given["node"]
+		_, module := a.given["module"]
+		_, secret := a.given["secret"]
+		if node || module || secret {
+			if err := need("node", "module", "secret"); err != nil {
+				return nil, fmt.Errorf("%w: a module's own secret is named by node, module and secret together", err)
+			}
+			return []string{"secret", "rotate", str("node"), str("module"), str("secret")}, nil
+		}
+		// Neither shape: the command says its usage, which names both, and that is the answer the
+		// caller needs.
+		return []string{"rotate"}, nil
+	case "settings":
+		// `settings set|clear` at a shell (novox/hq issue 198). The values travel as an argument
+		// because a tool has no file to hand the command; the command reads either.
+		if err := need("module"); err != nil {
+			return nil, err
+		}
+		var argv []string
+		if on("clear") {
+			argv = []string{"settings", "clear", str("module")}
+		} else {
+			argv = []string{"settings", "set", str("module")}
+			// Neither values nor clear: the command says its usage, which names both.
+			if v := str("values"); v != "" {
+				argv = append(argv, v)
+			}
+		}
+		if n := str("node"); n != "" {
+			argv = append(argv, "--node", n)
+		}
+		return argv, nil
+	case "issue":
+		// The same act as `module issue` at a shell (novox/hq design 25 §4): the account is minted
+		// into the mesh's records and delivered at the machine's next push, which is the caller's to
+		// ask for — so the mesh is never pushed as a side effect of a credential.
+		if err := need("node", "module"); err != nil {
+			return nil, err
+		}
+		return []string{"module", "issue", str("module"), "--node", str("node")}, nil
+	case "build":
+		if err := need("repository"); err != nil {
+			return nil, err
+		}
+		// Not waited for: a tool call cannot hold a connection for the minutes a build takes; the
+		// daemon takes the outcome in when it comes and the id follows the build (issue 176). A
+		// repository given without a scheme is a path on the forge holding the git seat.
+		argv := []string{"build", str("repository"), "--wait", "0"}
+		if !strings.Contains(str("repository"), "://") && !strings.HasPrefix(str("repository"), "git@") {
+			argv = append(argv, "--self")
+		}
+		if p := str("path"); p != "" {
+			argv = append(argv, "--path", p)
+		}
+		if r := str("ref"); r != "" {
+			argv = append(argv, "--ref", r)
+		}
+		return argv, nil
+	}
+	return nil, fmt.Errorf("%q is a verb of the %s seat's table that this binary has no command line for",
+		verb, catalogue.ControllerSeatName)
+}
+
+// jsonVerbs are the verbs whose command speaks JSON, so the answer carries it as data as well.
+var jsonVerbs = map[string]bool{"status": true, "seats": true, "plan": true, "collection": true}
+
+// runVerb runs this binary with the given command line and gathers what it said.
+func runVerb(ctx context.Context, argv []string) (verbAnswer, error) {
+	self, err := os.Executable()
+	if err != nil {
+		return verbAnswer{}, err
+	}
+	cmd := exec.CommandContext(ctx, self, argv...)
+	// The same environment: the stores' credentials, the bus, the broker — everything a command run
+	// from a shell in this container would have, because it is that.
+	cmd.Env = os.Environ()
+	// Two buffers, one answer. What the command *says* is both streams, in the order a person at
+	// a shell would read them; what it *answers as data* is standard output alone — `status --json`
+	// prints its warnings beside the document, and a JSON parsed from the two together parsed
+	// nothing (2026-09-30, the first status asked through the console had no `answer`).
+	var stdout, stderr bytes.Buffer
+	cmd.Stdout = &stdout
+	cmd.Stderr = &stderr
+	runErr := cmd.Run()
+	answer := verbAnswer{Output: stdout.String() + stderr.String(), OK: runErr == nil}
+	if jsonVerbs[argv[0]] && runErr == nil {
+		var parsed any
+		if json.Unmarshal(bytes.TrimSpace(stdout.Bytes()), &parsed) == nil {
+			answer.Answer = parsed
+		}
+	}
+	var exit *exec.ExitError
+	if runErr != nil && !errors.As(runErr, &exit) {
+		// Not the command refusing — the command not running at all, which is this process's fault.
+		return answer, fmt.Errorf("could not run %s: %w", strings.Join(argv, " "), runErr)
+	}
+	return answer, nil
+}
+
+// seatToolHandlers are the handlers for every verb the mesh-controller seat declares, from the
+// store's row, so a verb the row does not carry is not served. A verb it carries that this binary
+// cannot run is named at start and answers the reason when called — never a refusal to serve, which
+// would take the whole control plane down for one word (novox/hq ADR 0185).
+func seatToolHandlers() (map[string]link.ToolHandler, []string, error) {
+	seat, known := catalogue.SeatNamed(catalogue.ControllerSeatName)
+	if !known {
+		return nil, nil, fmt.Errorf("this mesh defines no %s seat", catalogue.ControllerSeatName)
+	}
+	var behind []string
+	handlers := map[string]link.ToolHandler{}
+	for _, v := range seat.Serves {
+		verb := v.Name
+		if inProcess[verb] {
+			handlers[verb] = func(ctx context.Context, raw json.RawMessage) (any, error) {
+				args := map[string]any{}
+				if len(bytes.TrimSpace(raw)) > 0 {
+					if err := json.Unmarshal(raw, &args); err != nil {
+						return nil, fmt.Errorf("the arguments are not a JSON object: %w", err)
+					}
+				}
+				// Refused like any verb's: what a verb does not take is not ignored.
+				a, err := readArguments(verb, args)
+				if err != nil {
+					return nil, err
+				}
+				if verb == "calls" {
+					return callsAnswer(link.Calls, a.given["call"])
+				}
+				return seatTools(), nil
+			}
+			continue
+		}
+		if _, err := argvFor(verb, sampleArguments(v)); err != nil {
+			// **A row ahead of this binary is not a reason to go silent.**
+			//
+			// The row is the store's and a control plane follows it (novox/hq ADR 0154), so a verb
+			// this build does not know means the row was widened by a newer one — the ordinary
+			// state of a roll-out, and of a push that put an older control plane back. Refusing to
+			// serve at all made that transient fatal: on 2026-10-02 one unknown verb took the whole
+			// mesh off the bus for ten minutes, and the way back was a human running the binary by
+			// hand, because the thing that would have repaired it is the thing that was down
+			// (novox/hq 04-ISSUES/201, ADR 0185).
+			//
+			// So the verbs this binary knows are served, and this one answers the reason instead of
+			// nothing: a caller gets a sentence naming the fault, and everything else keeps working
+			// — including the push that replaces this binary with the one whose verb it is.
+			behind = append(behind, verb)
+			reason := err
+			handlers[verb] = func(context.Context, json.RawMessage) (any, error) {
+				return nil, fmt.Errorf("%s is in this mesh's %s row and the control plane running "+
+					"here cannot run it: %w. It is a verb of a newer build; this one is behind",
+					verb, catalogue.ControllerSeatName, reason)
+			}
+			continue
+		}
+		handlers[verb] = func(ctx context.Context, raw json.RawMessage) (any, error) {
+			args := map[string]any{}
+			if len(raw) > 0 {
+				if err := json.Unmarshal(raw, &args); err != nil {
+					return nil, fmt.Errorf("the arguments are not a JSON object: %w", err)
+				}
+			}
+			argv, err := argvFor(verb, args)
+			if err != nil {
+				return nil, err
+			}
+			if answersFirst(argv) {
+				// Before anything is sent: a push sends the bus's own machine first, and a broker
+				// reloading its user list forgets the answer it was about to permit (novox/hq issue 265).
+				link.Acknowledge(ctx)
+			}
+			return runVerb(ctx, argv)
+		}
+	}
+	return handlers, behind, nil
+}
+
+// inProcess are the verbs answered by this process rather than by a command it runs: `tools` from
+// the records, `calls` from what this process served.
+var inProcess = map[string]bool{"tools": true, "calls": true}
+
+// answersFirst is a command line whose caller is answered before it runs: a push, by its verb or
+// through `command`. A push sends the machine holding the bus first when its user list changed, the
+// broker reloads, and a reload forgets every answer the bus was about to permit — so an answer
+// waiting for the push to end was refused, every time the list had changed (novox/hq issue 265).
+func answersFirst(argv []string) bool {
+	return len(argv) > 0 && argv[0] == "push"
+}
+
+// callsAnswer is what `calls` answers: the kept calls, newest first, without their answers — or
+// one call whole.
+func callsAnswer(log *link.CallLog, id string) (any, error) {
+	if id != "" {
+		c, ok := log.Get(id)
+		if !ok {
+			return nil, fmt.Errorf("no call %s is kept here: calls are kept by the controller that "+
+				"answered them, the last %d, and not across a restart — `calls` lists them", id, link.KeptCalls)
+		}
+		return c, nil
+	}
+	recent := log.Recent()
+	for i := range recent {
+		recent[i].Answer = nil
+	}
+	return map[string]any{"calls": recent, "kept": link.KeptCalls,
+		"note": "newest first; `calls` with a call's id gives its whole answer"}, nil
+}
+
+// seatTools is what `tools` answers: every seat with a protocol, and the tools each serves, from the
+// mesh's own records — no holder in the path, so it is true while a holder restarts (design 33 §5).
+func seatTools() map[string]any {
+	var seats []map[string]any
+	for _, s := range catalogue.SeatsWithAProtocol() {
+		if len(s.Serves) == 0 {
+			continue
+		}
+		var tools []map[string]any
+		for _, v := range s.Serves {
+			tools = append(tools, map[string]any{
+				"name": v.Name, "description": v.Description, "input": v.Input, "output": v.Output,
+			})
+		}
+		seats = append(seats, map[string]any{"seat": s.Name, "scope": s.Scope, "tools": tools})
+	}
+	return map[string]any{"seats": seats}
+}
+
+// sampleArguments is one of every argument a verb's schema requires, so the check at start proves the
+// verb runnable rather than that it happens to want the arguments the check guessed — and nothing
+// more, since an argument a verb does not declare is refused.
+func sampleArguments(v catalogue.Verb) map[string]any {
+	sample := map[string]any{}
+	switch required := v.Input["required"].(type) {
+	case []string:
+		for _, k := range required {
+			sample[k] = "x"
+		}
+	case []any:
+		for _, k := range required {
+			if name, ok := k.(string); ok {
+				sample[name] = "x"
+			}
+		}
+	}
+	return sample
+}
+
+// splitCommandLine splits a command line into words the way a POSIX shell does for the simple
+// cases a controller command needs: spaces separate, single or double quotes group, a backslash
+// escapes the next character inside double quotes or outside any. No expansion of anything.
+func splitCommandLine(line string) ([]string, error) {
+	var words []string
+	var cur strings.Builder
+	inWord := false
+	quote := rune(0)
+	runes := []rune(line)
+	for i := 0; i < len(runes); i++ {
+		r := runes[i]
+		switch {
+		case quote == '\'':
+			if r == '\'' {
+				quote = 0
+			} else {
+				cur.WriteRune(r)
+			}
+		case quote == '"':
+			if r == '"' {
+				quote = 0
+			} else if r == '\\' && i+1 < len(runes) {
+				i++
+				cur.WriteRune(runes[i])
+			} else {
+				cur.WriteRune(r)
+			}
+		case r == '\'' || r == '"':
+			quote = r
+			inWord = true
+		case r == '\\' && i+1 < len(runes):
+			i++
+			cur.WriteRune(runes[i])
+			inWord = true
+		case r == ' ' || r == '\t' || r == '\n':
+			if inWord {
+				words = append(words, cur.String())
+				cur.Reset()
+				inWord = false
+			}
+		default:
+			cur.WriteRune(r)
+			inWord = true
+		}
+	}
+	if quote != 0 {
+		return nil, fmt.Errorf("command has an unclosed %c quote", quote)
+	}
+	if inWord {
+		words = append(words, cur.String())
+	}
+	return words, nil
+}
+
+// seatAnnouncement is what the controller says it serves on the bus (novox/hq ADR 0197): the
+// mesh-controller seat, one endpoint per verb it answers, each with the seat's own description and
+// argument schema — the same facts `tools` answers from the records, as NATS's services format.
+func seatAnnouncement(handlers map[string]link.ToolHandler) micro.Info {
+	about := map[string]catalogue.Verb{}
+	for _, s := range catalogue.SeatsWithAProtocol() {
+		if s.Name == catalogue.ControllerSeatName {
+			for _, v := range s.Serves {
+				about[v.Name] = v
+			}
+		}
+	}
+	verbs := make([]string, 0, len(handlers))
+	for verb := range handlers {
+		verbs = append(verbs, verb)
+	}
+	sort.Strings(verbs)
+	var endpoints []micro.EndpointInfo
+	for _, verb := range verbs {
+		schema, _ := json.Marshal(about[verb].Input)
+		// The same shape every tool runtime announces in (node-tools' announce package): the name is
+		// `<seat>__<verb>`, as the protocol's characters allow; the metadata is what identifies it.
+		endpoints = append(endpoints, micro.EndpointInfo{
+			Name:       catalogue.ControllerSeatName + "__" + verb,
+			Subject:    link.SeatToolSubject(catalogue.ControllerSeatName, verb),
+			QueueGroup: "seat." + catalogue.ControllerSeatName,
+			Metadata: map[string]string{
+				"kind": "seat", "module": catalogue.ControllerSeatName, "tool": verb,
+				"seat": catalogue.ControllerSeatName, "scope": "mesh", "interchangeable": "false",
+				"description": about[verb].Description, "schema": string(schema),
+			},
+		})
+	}
+	return micro.Info{
+		ServiceIdentity: micro.ServiceIdentity{
+			Name: catalogue.ControllerSeatName, ID: "controller", Version: "0.1.0",
+			Metadata: map[string]string{"seat": catalogue.ControllerSeatName, "scope": "mesh"},
+		},
+		Description: "the mesh's own verbs, answered by the holder of the mesh-controller seat",
+		Endpoints:   endpoints,
+	}
+}

cmd/mesh-controller/seatverbs_schema_test.go

@@ -0,0 +1,365 @@
+package main
+
+import (
+	"encoding/json"
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"path/filepath"
+	"reflect"
+	"sort"
+	"strconv"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// The schemas the controller serves, verb by verb, as the console receives them: from the
+// announcement, not the table — what is checked is what a caller is shown (novox/hq issue 244).
+func servedSchemas(t *testing.T) map[string]catalogue.Verb {
+	t.Helper()
+	handlers, behind, err := seatToolHandlers()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(behind) != 0 {
+		t.Fatalf("this build cannot run %v of its own seat's verbs", behind)
+	}
+	served := map[string]catalogue.Verb{}
+	for _, e := range seatAnnouncement(handlers).Endpoints {
+		var input map[string]any
+		if err := json.Unmarshal([]byte(e.Metadata["schema"]), &input); err != nil {
+			t.Fatalf("%s announces a schema that is not JSON: %v", e.Name, err)
+		}
+		served[e.Metadata["tool"]] = catalogue.Verb{Name: e.Metadata["tool"], Input: input}
+	}
+	if len(served) != len(catalogue.ControllerVerbs) {
+		t.Fatalf("%d verbs served for %d in the table", len(served), len(catalogue.ControllerVerbs))
+	}
+	return served
+}
+
+// subsetsOf is every subset of the names, the empty one included.
+func subsetsOf(names []string) [][]string {
+	var out [][]string
+	for mask := 0; mask < 1<<len(names); mask++ {
+		var s []string
+		for i, n := range names {
+			if mask&(1<<i) != 0 {
+				s = append(s, n)
+			}
+		}
+		out = append(out, s)
+	}
+	return out
+}
+
+func argumentsFor(subset []string, switches map[string]bool) map[string]any {
+	args := map[string]any{}
+	for _, n := range subset {
+		if switches[n] {
+			args[n] = "true"
+		} else {
+			args[n] = "x-" + n
+		}
+	}
+	return args
+}
+
+// saidOutright are the switches that say the default aloud, so the line is the same without them —
+// on purpose, and only these.
+var saidOutright = map[string]string{
+	"push": "behind", // the whole mesh is what no machine means; behind lets a caller say they meant it
+}
+
+// **No argument a caller gives is passed over** (novox/hq issue 244). For every verb served and
+// every combination of the arguments its schema declares, the verb either refuses the call, or
+// composes a command line that each given argument changed: taking any one away changes the line
+// or makes it refused. An argument the line is the same without is one the verb ignored — the
+// shape of the push that named a machine and pushed every machine behind.
+func TestNoArgumentAVerbIsGivenIsPassedOver(t *testing.T) {
+	for name, v := range servedSchemas(t) {
+		if inProcess[name] {
+			continue
+		}
+		names, switches := declaredArguments(v)
+		for _, subset := range subsetsOf(names) {
+			args := argumentsFor(subset, switches)
+			argv, err := argvFor(name, args)
+			if err != nil {
+				if strings.Contains(err.Error(), "does not declare") {
+					t.Errorf("%s %v: %v", name, args, err)
+				}
+				continue
+			}
+			for _, dropped := range subset {
+				fewer := map[string]any{}
+				for k, val := range args {
+					if k != dropped {
+						fewer[k] = val
+					}
+				}
+				without, err := argvFor(name, fewer)
+				if err == nil && reflect.DeepEqual(argv, without) && saidOutright[name] != dropped {
+					t.Errorf("%s ignores %q given with %v: %v either way", name, dropped, subset, argv)
+				}
+			}
+		}
+	}
+}
+
+// **An argument a verb does not declare is refused, naming it — never dropped.** Every verb, with
+// what it requires and one argument more; and `node` in particular, for every verb whose schema
+// does not take a machine.
+func TestAnArgumentAVerbDoesNotDeclareIsRefused(t *testing.T) {
+	for name, v := range servedSchemas(t) {
+		names, _ := declaredArguments(v)
+		strangers := []string{"no-such-argument"}
+		if !contains(names, "node") {
+			strangers = append(strangers, "node")
+		}
+		for _, stranger := range strangers {
+			args := sampleArguments(v)
+			args[stranger] = "x"
+			_, err := argvFor(name, args)
+			if inProcess[name] {
+				_, err = readArguments(name, args)
+			}
+			if err == nil || !strings.Contains(err.Error(), strconv.Quote(stranger)) {
+				t.Errorf("%s took %q, which it does not declare: %v", name, stranger, err)
+			}
+		}
+	}
+	if _, err := argvFor("clear", map[string]any{"dead": "yes"}); err == nil {
+		t.Error("a switch took a word that is neither true nor false, and would have read it as false")
+	}
+	if _, err := argvFor("node", map[string]any{"node": 7}); err == nil {
+		t.Error("a number was taken as a machine's name, or as no machine")
+	}
+	if argv, err := argvFor("clear", map[string]any{"dead": true}); err != nil || strings.Join(argv, " ") != "clear --dead" {
+		t.Errorf("a switch given as JSON true: %v %v", argv, err)
+	}
+}
+
+// The push that was the cause: a machine named is that machine; none named is the whole mesh, and
+// naming one beside behind is refused rather than one of the two guessed.
+func TestAPushIsOneMachineOrSaysItIsTheWholeMesh(t *testing.T) {
+	argv, err := argvFor("push", map[string]any{"node": "g1"})
+	if err != nil || strings.Join(argv, " ") != "push g1 --wait 0" {
+		t.Fatalf("a named push: %v %v", argv, err)
+	}
+	for _, args := range []map[string]any{{}, {"behind": "true"}} {
+		argv, err := argvFor("push", args)
+		if err != nil || strings.Join(argv, " ") != "push --behind --wait 0" {
+			t.Fatalf("a push of the whole mesh %v: %v %v", args, argv, err)
+		}
+	}
+	if _, err := argvFor("push", map[string]any{"node": "g1", "behind": "true"}); err == nil ||
+		!strings.Contains(err.Error(), `"behind"`) {
+		t.Fatalf("a named push with behind was taken: %v", err)
+	}
+	if _, err := argvFor("push", map[string]any{"machine": "g1"}); err == nil || !strings.Contains(err.Error(), `"machine"`) {
+		t.Fatalf("a push given the machine under another name ran as a push of every machine: %v", err)
+	}
+}
+
+// commandFlags is every flag set this package's commands parse, by the name the set is made with,
+// and the flags defined on it — read from the source, so a flag added to a command is seen here
+// without anyone remembering to.
+func commandFlags(t *testing.T) map[string][]string {
+	t.Helper()
+	files, err := filepath.Glob("*.go")
+	if err != nil {
+		t.Fatal(err)
+	}
+	fset := token.NewFileSet()
+	out := map[string][]string{}
+	for _, f := range files {
+		if strings.HasSuffix(f, "_test.go") {
+			continue
+		}
+		file, err := parser.ParseFile(fset, f, nil, 0)
+		if err != nil {
+			t.Fatal(err)
+		}
+		for _, decl := range file.Decls {
+			fn, ok := decl.(*ast.FuncDecl)
+			if !ok || fn.Body == nil {
+				continue
+			}
+			sets := map[string]string{} // variable → the set's name
+			ast.Inspect(fn.Body, func(n ast.Node) bool {
+				if assign, ok := n.(*ast.AssignStmt); ok && len(assign.Lhs) == 1 && len(assign.Rhs) == 1 {
+					if call, ok := assign.Rhs[0].(*ast.CallExpr); ok && isSelector(call.Fun, "flag", "NewFlagSet") {
+						if id, ok := assign.Lhs[0].(*ast.Ident); ok {
+							if name, ok := stringLit(call.Args[0]); ok {
+								sets[id.Name] = name
+								out[name] = append(out[name], []string{}...)
+							}
+						}
+					}
+				}
+				call, ok := n.(*ast.CallExpr)
+				if !ok {
+					return true
+				}
+				sel, ok := call.Fun.(*ast.SelectorExpr)
+				if !ok {
+					return true
+				}
+				recv, ok := sel.X.(*ast.Ident)
+				if !ok || sets[recv.Name] == "" {
+					return true
+				}
+				arg := 0
+				switch sel.Sel.Name {
+				case "Bool", "String", "Int", "Int64", "Uint", "Uint64", "Float64", "Duration", "Func", "BoolFunc", "TextVar":
+				case "BoolVar", "StringVar", "IntVar", "Int64Var", "UintVar", "Uint64Var", "Float64Var", "DurationVar", "Var":
+					arg = 1
+				default:
+					return true
+				}
+				if arg < len(call.Args) {
+					if flagName, ok := stringLit(call.Args[arg]); ok {
+						out[sets[recv.Name]] = append(out[sets[recv.Name]], flagName)
+					}
+				}
+				return true
+			})
+		}
+	}
+	return out
+}
+
+func isSelector(e ast.Expr, pkg, name string) bool {
+	sel, ok := e.(*ast.SelectorExpr)
+	if !ok {
+		return false
+	}
+	id, ok := sel.X.(*ast.Ident)
+	return ok && id.Name == pkg && sel.Sel.Name == name
+}
+
+func stringLit(e ast.Expr) (string, bool) {
+	lit, ok := e.(*ast.BasicLit)
+	if !ok || lit.Kind != token.STRING {
+		return "", false
+	}
+	s, err := strconv.Unquote(lit.Value)
+	return s, err == nil
+}
+
+// accountedFlags are the flags of a verb's command that are not an argument of the same name:
+// carried by an argument named otherwise ("=argument"), or set by the verb itself, or withheld from
+// the named verb on purpose — each with why. `command` reaches every flag of the binary regardless.
+var accountedFlags = map[string]map[string]string{
+	"status": {"json": "set by the verb: the answer is data"},
+	"seats":  {"json": "set by the verb: the answer is data"},
+	"queue":  {"json": "not set: the verb answers the table a person reads"},
+	"plan":   {"json": "set by the verb unless files is asked"},
+	"push":   {"wait": "set by the verb to 0: a tool call cannot hold a connection for a whole apply"},
+	"build": {
+		"wait":    "set by the verb to 0: the id follows the build (issue 176)",
+		"self":    "set by the verb from the repository's form: a path on the forge, or a URL",
+		"dry-run": "withheld: a dry run answers only when the build ends, which a call cannot wait for; `command` reaches it",
+		"behind":  "withheld: the named verb builds one named repository; `command` reaches the rest",
+		"on":      "withheld: the named verb builds one named repository; `command` reaches the rest",
+	},
+	"builds": {"n": "=limit"},
+	"plans":  {"n": "=limit", "what-if": "=repository"},
+}
+
+// **Every flag of the command a verb runs is in the verb's schema, or accounted for here.** Derived
+// from the source, so a flag added to a command — or a verb added whose command takes flags —
+// fails this until somebody decides, in writing, how a caller reaches it (novox/hq issue 244). And
+// a flag accounted for that no longer exists fails too, so the table cannot rot into a list nobody
+// reads.
+func TestEveryFlagOfAVerbsCommandIsAnArgumentOrAccountedFor(t *testing.T) {
+	flags := commandFlags(t)
+	if len(flags["push"]) == 0 || len(flags["plan"]) == 0 {
+		t.Fatalf("reading the commands' flags found nothing for push or plan: %v", flags)
+	}
+	reached := map[string]map[string]bool{} // verb → flag sets its command lines reach
+	served := servedSchemas(t)
+	for name, v := range served {
+		if inProcess[name] || name == "command" {
+			continue
+		}
+		names, switches := declaredArguments(v)
+		for _, subset := range subsetsOf(names) {
+			argv, err := argvFor(name, argumentsFor(subset, switches))
+			if err != nil {
+				continue
+			}
+			// The flag set is named by the longest run of leading words that names one.
+			var words []string
+			for _, w := range argv {
+				if strings.HasPrefix(w, "-") {
+					break
+				}
+				words = append(words, w)
+			}
+			for n := len(words); n > 0; n-- {
+				if _, has := flags[strings.Join(words[:n], " ")]; has {
+					if reached[name] == nil {
+						reached[name] = map[string]bool{}
+					}
+					reached[name][strings.Join(words[:n], " ")] = true
+					break
+				}
+			}
+		}
+	}
+	used := map[string]map[string]bool{}
+	verbs := make([]string, 0, len(reached))
+	for verb := range reached {
+		verbs = append(verbs, verb)
+	}
+	sort.Strings(verbs)
+	for _, verb := range verbs {
+		declared, _ := declaredArguments(served[verb])
+		for set := range reached[verb] {
+			if used[set] == nil {
+				used[set] = map[string]bool{}
+			}
+			for _, flagName := range flags[set] {
+				why, accounted := accountedFlags[set][flagName]
+				switch {
+				case accounted && strings.HasPrefix(why, "="):
+					used[set][flagName] = true
+					if !contains(declared, strings.TrimPrefix(why, "=")) {
+						t.Errorf("%s: --%s of `%s` is said to be carried by %q, which the schema does not declare",
+							verb, flagName, set, strings.TrimPrefix(why, "="))
+					}
+				case accounted:
+					used[set][flagName] = true
+				case contains(declared, flagName):
+				default:
+					t.Errorf("%s runs `%s`, which takes --%s, and the verb's schema has no %q: declare it, "+
+						"or say in accountedFlags why a caller does not reach it", verb, set, flagName, flagName)
+				}
+			}
+		}
+	}
+	for set, fs := range accountedFlags {
+		for flagName := range fs {
+			if !used[set][flagName] {
+				t.Errorf("accountedFlags names --%s of `%s`, which no verb's command takes any more", flagName, set)
+			}
+		}
+	}
+}
+
+// Every verb's required arguments are properties of its schema: a schema that requires what it
+// does not describe is the uncallable verb of issue 244 from the other side.
+func TestEveryRequiredArgumentIsDescribed(t *testing.T) {
+	for name, v := range servedSchemas(t) {
+		names, _ := declaredArguments(v)
+		for k := range sampleArguments(v) {
+			if !contains(names, k) {
+				t.Errorf("%s requires %q and does not describe it", name, k)
+			}
+		}
+	}
+}

cmd/mesh-controller/seatverbs_test.go

@@ -0,0 +1,280 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"github.com/novox/mesh-controller/internal/link"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// `builds` given a build's id reads that build's log from the bus rather than listing builds
+// (novox/hq ADR 0157).
+func TestBuildsWithAnIdReadsThatBuildsLog(t *testing.T) {
+	argv, err := argvFor("builds", map[string]any{"log": "build-17"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Join(argv, " ") != "builds --log build-17" {
+		t.Fatalf("builds with a log id became %q", strings.Join(argv, " "))
+	}
+}
+
+// The build tool takes a repository as a URL or as its path on the forge holding the git seat, and
+// says which it was given, so the command reads the path as a seat source rather than handing it to
+// git as written (novox/hq issue 176). And it never waits: the id follows the build.
+func TestTheBuildToolTellsAForgePathFromAURL(t *testing.T) {
+	argv, _ := argvFor("build", map[string]any{"repository": "novox/mesh-catalog", "path": "modules/x"})
+	if line := strings.Join(argv, " "); !strings.Contains(line, "--self") || !strings.Contains(line, "--wait 0") {
+		t.Fatalf("a forge path is a seat source, not waited for; got %q", line)
+	}
+	argv, _ = argvFor("build", map[string]any{"repository": "https://example.tld/o/r.git"})
+	if line := strings.Join(argv, " "); strings.Contains(line, "--self") {
+		t.Fatalf("a URL is cloned as given; got %q", line)
+	}
+}
+
+// `rotate` is one verb with two shapes (ADR 0114, issue 180): a pair credential by provision, or a
+// module's own secret by machine, module and name.
+func TestRotateTakesAProvisionOrAnOwnSecret(t *testing.T) {
+	argv, _ := argvFor("rotate", map[string]any{"provision": "postgres-database", "consumer": "ace"})
+	if strings.Join(argv, " ") != "rotate postgres-database --consumer ace" {
+		t.Fatalf("a pair credential: %v", argv)
+	}
+	argv, _ = argvFor("rotate", map[string]any{"node": "ace", "module": "nodered", "secret": "api-token"})
+	if strings.Join(argv, " ") != "secret rotate ace nodered api-token" {
+		t.Fatalf("an own secret: %v", argv)
+	}
+	if _, err := argvFor("rotate", map[string]any{"node": "ace"}); err == nil || !strings.Contains(err.Error(), `"module"`) {
+		t.Fatalf("half an own secret is refused, naming what it lacks: %v", err)
+	}
+	if argv, _ := argvFor("rotate", nil); strings.Join(argv, " ") != "rotate" {
+		t.Fatalf("neither shape falls to the command's usage: %v", argv)
+	}
+	if _, err := argvFor("rotate", map[string]any{"provision": "p", "node": "ace", "module": "m", "secret": "s"}); err == nil {
+		t.Fatal("both shapes at once were taken, and one of them passed over")
+	}
+}
+
+// `settings` is `settings set|clear` at a shell, with the values passed inline (novox/hq issue 198).
+func TestSettingsSetsOrClearsALayer(t *testing.T) {
+	argv, err := argvFor("settings", map[string]any{"module": "dnsmasq", "values": `{"a":1}`, "node": "ace"})
+	if err != nil || strings.Join(argv, " ") != `settings set dnsmasq {"a":1} --node ace` {
+		t.Fatalf("set on a machine: %v %v", argv, err)
+	}
+	argv, _ = argvFor("settings", map[string]any{"module": "dnsmasq", "clear": "true"})
+	if strings.Join(argv, " ") != "settings clear dnsmasq" {
+		t.Fatalf("clear for the mesh: %v", argv)
+	}
+	argv, _ = argvFor("settings", map[string]any{"module": "dnsmasq"})
+	if strings.Join(argv, " ") != "settings set dnsmasq" {
+		t.Fatalf("a set with no values falls to the command's usage: %v", argv)
+	}
+}
+
+// `issue` is `module issue` at a shell: the module and the machine, and nothing that would push. A
+// module's bus account was mintable only from the controller's command line, so an agent working
+// through the tools could not finish a rollout that gave a module one (novox/hq issue 191).
+func TestIssueGivesAModuleItsAccountOnAMachine(t *testing.T) {
+	argv, err := argvFor("issue", map[string]any{"node": "ace", "module": "route-proxy"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Join(argv, " ") != "module issue route-proxy --node ace" {
+		t.Fatalf("issue runs %v", argv)
+	}
+	if _, err := argvFor("issue", map[string]any{"module": "route-proxy"}); err == nil {
+		t.Error("an account was issued without saying which machine reads it")
+	}
+}
+
+// A required argument missing is refused in the verb's own words, before anything runs.
+func TestAVerbMissingWhatItNeedsIsRefused(t *testing.T) {
+	if _, err := argvFor("node", map[string]any{}); err == nil || !strings.Contains(err.Error(), `node needs "node"`) {
+		t.Fatalf("node without a machine was accepted: %v", err)
+	}
+	if _, err := argvFor("upgrade", map[string]any{}); err == nil {
+		t.Fatal("a verb the seat does not serve was accepted")
+	}
+}
+
+// A push and a build are sent, not waited for: the asker reads status, or the build's log by its
+// id, for what happened. A repository given as a forge path is said to be one (issue 176).
+func TestActsDoNotBlockTheCall(t *testing.T) {
+	argv, _ := argvFor("push", map[string]any{"node": "one"})
+	if strings.Join(argv, " ") != "push one --wait 0" {
+		t.Fatalf("push waits: %v", argv)
+	}
+	argv, _ = argvFor("build", map[string]any{"repository": "novox/x", "path": "modules/x"})
+	if strings.Join(argv, " ") != "build novox/x --wait 0 --self --path modules/x" {
+		t.Fatalf("build: %v", argv)
+	}
+}
+
+// What `tools` answers is the seats' records, with each verb's schema.
+func TestToolsAnswersTheSeatsRecords(t *testing.T) {
+	handlers, behind, err := seatToolHandlers()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(behind) != 0 {
+		t.Fatalf("this build cannot run %v of its own seat's verbs", behind)
+	}
+	if len(handlers) != len(catalogue.ControllerVerbs) {
+		t.Fatalf("%d handlers for %d verbs", len(handlers), len(catalogue.ControllerVerbs))
+	}
+	answer := seatTools()
+	seats, _ := answer["seats"].([]map[string]any)
+	var found bool
+	for _, s := range seats {
+		if s["seat"] == catalogue.ControllerSeatName {
+			found = true
+			tools, _ := s["tools"].([]map[string]any)
+			if len(tools) != len(catalogue.ControllerVerbs) || tools[0]["input"] == nil {
+				t.Fatalf("the controller seat's tools are not listed in full: %v", tools)
+			}
+		}
+	}
+	if !found {
+		t.Fatal("the mesh-controller seat is not in the listing")
+	}
+}
+
+// A JSON verb's answer is parsed from what the command wrote to standard output alone; a warning it
+// printed beside the document does not take the document away. The test binary stands in for the
+// controller: `-test.run` with a name that matches nothing prints `ok` and a warning about no tests.
+func TestAJSONVerbsAnswerIsItsStandardOutput(t *testing.T) {
+	jsonVerbs["-test.run"] = true
+	t.Cleanup(func() { delete(jsonVerbs, "-test.run") })
+	answer, err := runVerb(t.Context(), []string{"-test.run", "TestAnswerEcho", "-test.v"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !answer.OK {
+		t.Fatalf("the command failed: %s", answer.Output)
+	}
+	if !strings.Contains(answer.Output, "PASS") {
+		t.Fatalf("stderr and stdout are both what the command said: %s", answer.Output)
+	}
+}
+
+// `command` is the generic verb: the command line as given, split as a shell would, nothing added —
+// so an operator's `node account g14 jochen` is one call through the console rather than a shell on
+// the control node (novox/hq ADR 0154, ADR 0175).
+func TestCommandRunsTheLineAsGiven(t *testing.T) {
+	argv, err := argvFor("command", map[string]any{"command": "node account g14 jochen"})
+	if err != nil || strings.Join(argv, " ") != "node account g14 jochen" {
+		t.Fatalf("a plain line: %v %v", argv, err)
+	}
+	argv, err = argvFor("command", map[string]any{"command": `settings set dnsmasq '{"a": "b c"}' --node ace`})
+	if err != nil || len(argv) != 6 || argv[3] != `{"a": "b c"}` {
+		t.Fatalf("a quoted word stays one word: %q %v", argv, err)
+	}
+	argv, err = argvFor("command", map[string]any{"command": `node add "the box" --adopted`})
+	if err != nil || len(argv) != 4 || argv[2] != "the box" {
+		t.Fatalf("double quotes group: %q %v", argv, err)
+	}
+	if _, err := argvFor("command", map[string]any{"command": "   "}); err == nil {
+		t.Fatal("an empty line was accepted")
+	}
+	if _, err := argvFor("command", map[string]any{"command": `node "unclosed`}); err == nil {
+		t.Fatal("an unclosed quote was accepted")
+	}
+}
+
+// A verb in the row that this binary cannot run does not take the control plane off the bus: the
+// rest are served, the unknown one answers the reason, and the start-up names it (novox/hq ADR
+// 0185). One unknown word cost the mesh ten minutes of silence on 2026-10-02, recoverable only by
+// a person running the binary by hand — the push that would have repaired it needs the control
+// plane that was down.
+func TestARowAheadOfThisBuildIsServedAnyway(t *testing.T) {
+	seat, known := catalogue.SeatNamed(catalogue.ControllerSeatName)
+	if !known {
+		t.Fatal("no controller seat")
+	}
+	// The row as a newer control plane would have written it: every verb this build knows, and one
+	// it does not.
+	widened := seat
+	widened.Serves = append(append([]catalogue.Verb{}, seat.Serves...),
+		catalogue.Verb{Name: "teleport", Description: "a verb from a build that does not exist yet"})
+	rows := catalogue.DefaultSeats()
+	for i := range rows {
+		if rows[i].Name == catalogue.ControllerSeatName {
+			rows[i] = widened
+		}
+	}
+	catalogue.UseSeats(rows)
+	t.Cleanup(func() { catalogue.UseSeats(catalogue.DefaultSeats()) })
+
+	handlers, behind, err := seatToolHandlers()
+	if err != nil {
+		t.Fatalf("a row with one unknown verb refused to serve at all: %v", err)
+	}
+	if len(behind) != 1 || behind[0] != "teleport" {
+		t.Fatalf("the verbs this build cannot run were reported as %v", behind)
+	}
+	if len(handlers) != len(widened.Serves) {
+		t.Fatalf("%d handlers for %d verbs in the row", len(handlers), len(widened.Serves))
+	}
+	for _, known := range []string{"status", "nodes", "push"} {
+		if handlers[known] == nil {
+			t.Errorf("%s is not served although this build knows it", known)
+		}
+	}
+	_, err = handlers["teleport"](context.Background(), nil)
+	if err == nil {
+		t.Fatal("the unknown verb answered as though it had run")
+	}
+	for _, want := range []string{"teleport", "cannot run it", "behind"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Errorf("the answer does not say %q: %v", want, err)
+		}
+	}
+}
+
+// novox/hq ADR 0195: the console's discovery reads the machines and the modules; they answer as JSON,
+// as status and seats do, so nothing parses a printed column.
+func TestTheNodesAndModulesVerbsAnswerAsJSON(t *testing.T) {
+	for verb, want := range map[string]string{"nodes": "[node list --json]", "modules": "[module list --json]"} {
+		argv, err := argvFor(verb, map[string]any{})
+		if err != nil {
+			t.Fatal(err)
+		}
+		if fmt.Sprint(argv) != want {
+			t.Errorf("%s runs %v, want %s", verb, argv, want)
+		}
+	}
+}
+
+// novox/hq ADR 0197: the controller announces exactly the verbs it serves, each on the subject and
+// queue it serves it on, with the seat's own description and schema, in NATS's services format.
+func TestTheControllerAnnouncesTheVerbsItServes(t *testing.T) {
+	handlers, _, err := seatToolHandlers()
+	if err != nil {
+		t.Fatal(err)
+	}
+	info := seatAnnouncement(handlers)
+	if info.Name != catalogue.ControllerSeatName || info.ID == "" || info.Version == "" {
+		t.Fatalf("the service is not named for the seat: %+v", info.ServiceIdentity)
+	}
+	if len(info.Endpoints) != len(handlers) {
+		t.Fatalf("%d endpoints announced for %d verbs served", len(info.Endpoints), len(handlers))
+	}
+	for _, e := range info.Endpoints {
+		verb := e.Metadata["tool"]
+		if _, served := handlers[verb]; !served || e.Name != catalogue.ControllerSeatName+"__"+verb {
+			t.Errorf("%s (%s) is announced and not served under that name", e.Name, verb)
+		}
+		if e.Metadata["kind"] != "seat" || e.Metadata["seat"] != catalogue.ControllerSeatName {
+			t.Errorf("%s is not announced as the seat's verb: %v", e.Name, e.Metadata)
+		}
+		if e.Subject != link.SeatToolSubject(catalogue.ControllerSeatName, verb) || e.QueueGroup != "seat."+catalogue.ControllerSeatName {
+			t.Errorf("%s is announced on %s/%s, not where it is served", e.Name, e.Subject, e.QueueGroup)
+		}
+		if e.Metadata["description"] == "" || e.Metadata["schema"] == "" || e.Metadata["scope"] != "mesh" {
+			t.Errorf("%s is announced without its description, schema or scope: %v", e.Name, e.Metadata)
+		}
+	}
+}

cmd/mesh-controller/secret.go

@@ -0,0 +1,421 @@
+package main
+
+import (
+	"bufio"
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"io"
+	"os"
+	"strings"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/secrets"
+)
+
+// secretCommand gives the mesh a value it must carry and could not have invented.
+//
+// **Every other secret in this mesh is one the mesh made** — generated, sealed to the machine that
+// will use it, and never readable again. That is right for something coming into existence, and
+// wrong for something that already exists: a database created last year has the password it was
+// created with, and generating a new one puts 32 random bytes where a working credential was.
+// The machine applies it, reports success, and whatever reads it fails to authenticate somewhere
+// else entirely — with the mesh insisting the secret was delivered, which it was.
+//
+// So this is the entry point for **adopting** something already running. The store has carried
+// the distinction since the beginning: a module secret records whether it was `made` or
+// `accepted`, and refuses to invent a replacement for the second. Nothing until now could write
+// one, so the only accepted secret in the mesh was the broker account issued to a build machine.
+//
+// The value is sealed on the way in and the plaintext discarded, exactly as a generated one is.
+// **The only difference between the two is where the value came from.**
+func secretCommand(ctx context.Context, args []string) error {
+	if len(args) == 0 {
+		return errors.New(secretUsage)
+	}
+	switch args[0] {
+	case "accept":
+	case "rotate":
+		return secretRotate(ctx, args[1:])
+	case "recover":
+		return secretRecover(ctx, args[1:])
+	case "export":
+		return secretExport(ctx, args[1:])
+	default:
+		return errors.New(secretUsage)
+	}
+	rest, flags := split(args[1:])
+	set := flag.NewFlagSet("secret accept", flag.ContinueOnError)
+	from := set.String("from", "",
+		"read the value from this file instead of asking (use - for standard input)")
+	provider := set.String("provider", "",
+		"the node providing <name>: the value becomes the PAIR credential between <module> on <node> "+
+			"and that provider, sealed to both — the vault's operator-delivered secret (ADR 0092)")
+	local := set.String("local", "",
+		"with --provider: the name the credential goes by inside <module>, where its manifest keeps "+
+			"several for <name> (ADR 0094)")
+	if err := set.Parse(flags); err != nil {
+		return err
+	}
+	if len(rest) != 3 {
+		return errors.New(secretUsage)
+	}
+	node, module, name := rest[0], rest[1], rest[2]
+
+	value, err := valueFor(node, module, name, *from)
+	if err != nil {
+		return err
+	}
+	value = asSupplied(value)
+	if value == "" {
+		return errors.New("there is nothing to seal")
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+
+	if *provider != "" {
+		// Into the pair, not into the module's own secrets: what the provider is asked to create
+		// and what the consumer reads are the same value, and neither end can be told a different
+		// one later without the other (novox/hq 04-ISSUES/070).
+		if err := open.inventory.AcceptSecretForPair(ctx, name, node, module, *provider, *local, value); err != nil {
+			return err
+		}
+		fmt.Printf("%s on %s now holds %q from %s%s, sealed to both machines.\n", module, node, name, *provider, asLocal(*local))
+		fmt.Printf("  the mesh cannot read it back, will not replace it with one of its own, and will not rotate it\n")
+		fmt.Printf("  run `push %s` and `push %s` to send it\n", *provider, node)
+		return nil
+	}
+	if err := open.inventory.AcceptSecretForModule(ctx, node, module, name, value); err != nil {
+		return err
+	}
+	// Not printed back, and there is nowhere it could be printed from: it is sealed to that
+	// machine and the mesh cannot read it again.
+	fmt.Printf("%s on %s now holds %q, sealed to that machine.\n", module, node, name)
+	fmt.Printf("  the mesh cannot read it back, and will not replace it with one of its own\n")
+	fmt.Printf("  run `push %s` to send it\n", node)
+	return nil
+}
+
+const secretUsage = "secret rotate <node> <module> <name>\n" +
+	"secret accept <node> <module> <name> [--from <file>] [--provider <node> [--local <name>]]\n" +
+	"secret recover <node> <module> <name> --key <operator-key> [--out <file>] [--from-export <file>] [--provider <node>]\n" +
+	"secret export [--out <file>]"
+
+// secretRecover is break-glass: a secret opened with the operator's key, written to a file.
+//
+// **The mesh cannot show a secret back, and this does not make it able to.** What is opened here
+// is the copy sealed to the operator key (novox/hq ADR 0085, amended); the mesh holds that blob and
+// no key for it, and this program holds the key for the length of the call and no blob until given
+// one. Recovery needs both, which is what keeps the sealing meaningful.
+//
+// The value goes to a file at 0600, never to the terminal unless asked for with `--out -` — the
+// source mesh's secret tools were written after a secret was printed into a transcript, and that
+// rule is theirs. `--from-export` reads the blob from a file `secret export` wrote, so recovery
+// works with the store gone, which is the case it exists for.
+func secretRecover(ctx context.Context, args []string) error {
+	rest, flags := split(args)
+	set := flag.NewFlagSet("secret recover", flag.ContinueOnError)
+	keyFile := set.String("key", "", "the operator's private key, from `operator key make`")
+	out := set.String("out", "", "where to write the value (0600); - for standard output. Default <node>.<module>.<name>.secret")
+	fromExport := set.String("from-export", "", "read the sealed copy from this `secret export` file instead of the store")
+	provider := set.String("provider", "", "for a pair credential held from more than one provider: which one")
+	local := set.String("local", "", "for a pair credential the module keeps under a local name (ADR 0094): which one")
+	if err := set.Parse(flags); err != nil {
+		return err
+	}
+	if len(rest) != 3 {
+		return errors.New(secretUsage)
+	}
+	node, module, name := rest[0], rest[1], rest[2]
+	private, err := readPrivateKey(*keyFile)
+	if err != nil {
+		return err
+	}
+
+	var kept inventory.Kept
+	if *fromExport != "" {
+		kept, err = keptFromExport(*fromExport, node, module, name, *provider)
+		if err != nil {
+			return err
+		}
+	} else {
+		open, err := openStores(ctx)
+		if err != nil {
+			return err
+		}
+		defer open.Close()
+		kept, err = open.inventory.KeptSecret(ctx, node, module, name, *provider, *local)
+		if err != nil {
+			return err
+		}
+	}
+
+	value, err := secrets.Open(private, kept.Sealed)
+	if err != nil {
+		return fmt.Errorf("%s on %s: %q is sealed to operator key %s, and that key does not open it: %w",
+			module, node, name, secrets.Fingerprint(kept.Key), err)
+	}
+	if *out == "-" {
+		_, err := os.Stdout.Write(append(value, '\n'))
+		return err
+	}
+	path := *out
+	if path == "" {
+		path = node + "." + module + "." + name + ".secret"
+	}
+	if err := writeNew(path, value); err != nil {
+		return err
+	}
+	fmt.Printf("%s on %s: %q recovered to %s (0600) — %d bytes, origin %s\n",
+		module, node, name, path, len(value), kept.Origin)
+	return nil
+}
+
+// An export is what a person keeps beside the operator key — the catalogue's shape, so the vault
+// keeps the same document on its disk (Manifest.Keeps).
+type export = catalogue.KeptExport
+
+func secretExport(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("secret export", flag.ContinueOnError)
+	out := set.String("out", "", "where to write the export (0600); - or empty for standard output")
+	if err := set.Parse(args); err != nil {
+		return err
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+	key, err := inv.OperatorKey(ctx)
+	if err != nil {
+		return err
+	}
+	if key == "" {
+		return errors.New("the mesh has no operator key, so nothing is sealed to one; `operator key make` and `operator key set` first")
+	}
+	doc, err := inv.OperatorExport(ctx)
+	if err != nil {
+		return err
+	}
+	body, err := json.MarshalIndent(doc, "", "  ")
+	if err != nil {
+		return err
+	}
+	body = append(body, '\n')
+	if *out == "" || *out == "-" {
+		_, err := os.Stdout.Write(body)
+		return err
+	}
+	// Replaced whole, and made 0600 whether or not it existed: an export is ciphertext and a public
+	// key, but it is also the list of every secret the mesh has, and a file left at an earlier mode
+	// while the command says 0600 is a lie in the one place a person checks.
+	if err := writeReplacing(*out, body); err != nil {
+		return err
+	}
+	fmt.Printf("%d secret(s) exported to %s (0600), sealed to operator key %s — ciphertext, keep it with the key\n",
+		len(doc.Kept), *out, doc.Fingerprint)
+	if len(doc.EarlierKey) > 0 {
+		fmt.Printf("  %d secret(s) are sealed to an EARLIER operator key: recoverable with that key only\n", len(doc.EarlierKey))
+	}
+	if len(doc.Unrecoverable) > 0 {
+		fmt.Printf("  %d secret(s) are NOT in it: made before the mesh had an operator key\n", len(doc.Unrecoverable))
+	}
+	return nil
+}
+
+// writeNew writes a file that must not exist yet, atomically: create-exclusive, 0600. A check
+// followed by a write is a window in which a key somebody still needs can be overwritten.
+func writeNew(path string, content []byte) error {
+	f, err := os.OpenFile(path, os.O_WRONLY|os.O_CREATE|os.O_EXCL, 0o600)
+	if err != nil {
+		if os.IsExist(err) {
+			return fmt.Errorf("%s already exists; not overwriting it", path)
+		}
+		return err
+	}
+	if _, err := f.Write(content); err != nil {
+		f.Close()
+		return err
+	}
+	return f.Close()
+}
+
+// writeReplacing writes a file whole, creating or truncating it, and leaves it at 0600 either way.
+func writeReplacing(path string, content []byte) error {
+	f, err := os.OpenFile(path, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o600)
+	if err != nil {
+		return err
+	}
+	if _, err := f.Write(content); err != nil {
+		f.Close()
+		return err
+	}
+	if err := f.Chmod(0o600); err != nil {
+		f.Close()
+		return err
+	}
+	return f.Close()
+}
+
+func keptFromExport(path, node, module, name, provider string) (inventory.Kept, error) {
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		return inventory.Kept{}, err
+	}
+	var e export
+	if err := json.Unmarshal(raw, &e); err != nil {
+		return inventory.Kept{}, fmt.Errorf("%s is not a secret export: %w", path, err)
+	}
+	// Sealed to the current key or to an earlier one: both are copies the given key might open,
+	// and Open says which. Not the unrecoverable list, which holds no copy at all.
+	var found []inventory.Kept
+	for _, k := range append(append([]inventory.Kept{}, e.Kept...), e.EarlierKey...) {
+		if k.Node == node && k.Module == module && k.Name == name && (provider == "" || k.Provider == provider) {
+			found = append(found, k)
+		}
+	}
+	switch len(found) {
+	case 0:
+		return inventory.Kept{}, fmt.Errorf("%s holds no copy of %s's %q on %s", path, module, name, node)
+	case 1:
+		return found[0], nil
+	default:
+		providers := make([]string, 0, len(found))
+		for _, f := range found {
+			providers = append(providers, f.Provider)
+		}
+		return inventory.Kept{}, fmt.Errorf("%s holds %s's %q on %s from more than one provider (%s); say which with --provider",
+			path, module, name, node, strings.Join(providers, ", "))
+	}
+}
+
+// split separates what this command is about from how it was asked.
+//
+// **Because the standard library stops parsing at the first non-flag argument.** With the
+// positionals first — which is the order that reads correctly — everything after them is left
+// sitting in the arguments, so `secret accept a b c --from -` arrives as five positionals and the
+// flag is never seen. The host's own parser carries the same note, and the fault it names is
+// worse than this one: there, a flag somebody passed was silently ignored and the command
+// succeeded anyway.
+func split(args []string) (positional, flags []string) {
+	for i, arg := range args {
+		if strings.HasPrefix(arg, "-") {
+			return args[:i], args[i:]
+		}
+	}
+	return args, nil
+}
+
+// asSupplied is the value with its line ending removed and nothing else.
+//
+// **A file has a trailing newline and a password does not**, so the ending goes — a credential
+// wrong by one byte fails in a way nobody connects to how it was supplied.
+//
+// **And only the ending.** Trimming both ends is the obvious thing and it is wrong: a password
+// chosen with a leading space is one the mesh would then deliver as a different password, silently,
+// with the operator certain they had supplied it correctly.
+func asSupplied(raw string) string {
+	return strings.TrimRight(raw, "\r\n")
+}
+
+// valueFor gets the secret without putting it somewhere it can be read afterwards.
+//
+// **Not an argument, and there is no flag that takes one.** A value on the command line is in the
+// shell's history, in the process list for as long as it runs, and in whatever collects either.
+// The paths here are a file the operator already has, or a prompt that does not echo — the same
+// two ways a model-access key is supplied (novox/hq ADR 0024).
+func valueFor(node, module, name, from string) (string, error) {
+	switch {
+	case from == "-":
+		body, err := io.ReadAll(os.Stdin)
+		if err != nil {
+			return "", err
+		}
+		return string(body), nil
+
+	case from != "":
+		body, err := os.ReadFile(from)
+		if err != nil {
+			return "", err
+		}
+		return string(body), nil
+
+	default:
+		// The same path a model-access key takes, and for the same reason: a value given as an
+		// argument is in the shell's history and in the process list. Read from standard input,
+		// echoed nowhere by this program.
+		fmt.Fprintf(os.Stderr,
+			"reading %s's %q for %s from standard input; it is not echoed anywhere\n",
+			module, name, node)
+		line, err := bufio.NewReader(os.Stdin).ReadString('\n')
+		if err != nil && line == "" {
+			return "", fmt.Errorf("nothing was given on standard input: %w", err)
+		}
+		return line, nil
+	}
+}
+
+// secretRotate makes a module's own secret anew and sends the machine, so the module starts again on
+// the new value (novox/hq ADR 0114, issue 180). A pair credential rotates with `rotate <provision>`;
+// this is the secret with one party. Said in the log with who asked and when, never the value.
+func secretRotate(ctx context.Context, args []string) error {
+	rest, _ := split(args)
+	if len(rest) != 3 {
+		return errors.New(secretUsage)
+	}
+	node, module, name := rest[0], rest[1], rest[2]
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	if err := open.inventory.RotateModuleSecret(ctx, node, module, name); err != nil {
+		var refused inventory.ErrNotRotatable
+		if errors.As(err, &refused) {
+			return fmt.Errorf("not rotated: %s", refused.Why)
+		}
+		return err
+	}
+	fmt.Printf("rotated %q of %s on %s at %s, asked by %s; the value is sealed and not shown\n",
+		name, module, node, time.Now().UTC().Format(time.RFC3339), whoAsked())
+	// A shared credential (ADR 0158) has as many holders as the provision has consumers, and all
+	// of them are sent in one act, so no machine is left reading a value the provider no longer takes.
+	machines, err := open.inventory.SharedHolders(ctx, node, module, name)
+	if err != nil {
+		return err
+	}
+	if len(machines) == 0 {
+		machines = []string{node}
+	}
+	if len(machines) == 1 {
+		fmt.Printf("sending %s, so %s starts again on the new value:\n", node, module)
+	} else {
+		fmt.Printf("shared with every consumer; sending %s together:\n", strings.Join(machines, ", "))
+	}
+	if err := sendTo(ctx, open, machines); err != nil {
+		return fmt.Errorf("%w\n\nThe new value is sealed and not yet delivered; what runs keeps the old "+
+			"one until the machines next apply. Fix the cause and run `push --behind`", err)
+	}
+	return nil
+}
+
+// whoAsked names the caller for the log: the account the command runs as, which for a tool call
+// through the console is the mesh's own.
+func whoAsked() string {
+	if u := os.Getenv("SUDO_USER"); u != "" {
+		return u
+	}
+	if u := os.Getenv("USER"); u != "" {
+		return u
+	}
+	return "the mesh"
+}

cmd/mesh-controller/secret_test.go

@@ -0,0 +1,111 @@
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// A file has a trailing newline and a password does not.
+//
+// The failure this prevents is the worst kind to diagnose: the credential is delivered, the
+// machine applies it, everything reports success, and authentication fails one byte from correct
+// somewhere else entirely.
+func TestATrailingNewlineIsNotPartOfTheSecret(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "password")
+	if err := os.WriteFile(path, []byte("the-database-password\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	got, err := valueFor("anchor", "umami", "database", path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if asSupplied(got) != "the-database-password" {
+		t.Fatalf("read %q", got)
+	}
+}
+
+// A password may contain spaces, and they are the operator's.
+//
+// Trimming both ends is the obvious thing and it is wrong: a value chosen with a leading space is
+// a value the mesh would silently deliver as a different one.
+func TestOnlyLineEndingsAreRemoved(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "password")
+	if err := os.WriteFile(path, []byte("  spaces matter  \n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	got, err := valueFor("anchor", "umami", "database", path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if asSupplied(got) != "  spaces matter  " {
+		t.Fatalf("the value was altered beyond its line ending: %q", got)
+	}
+}
+
+// A file that is not there is said plainly, rather than becoming an empty secret.
+func TestAMissingFileIsRefused(t *testing.T) {
+	if _, err := valueFor("anchor", "umami", "database",
+		filepath.Join(t.TempDir(), "absent")); err == nil {
+		t.Fatal("a missing file produced a value")
+	}
+}
+
+// The command refuses what it cannot act on, rather than acting on part of it.
+func TestTheArgumentsAreRequired(t *testing.T) {
+	for _, args := range [][]string{
+		{},
+		{"accept"},
+		{"accept", "anchor"},
+		{"accept", "anchor", "umami"},
+		{"give", "anchor", "umami", "database"},
+	} {
+		if err := secretCommand(t.Context(), args); err == nil {
+			t.Errorf("%v was accepted", args)
+		}
+	}
+}
+
+// The invocation that actually gets used, which the first version of these tests never tried.
+//
+// Every case here was a rejection, so the command was broken in the one way that matters — it
+// refused what it is for — and the tests were green. The lab found it at the first call.
+func TestTheArgumentsAndTheFlagAreBothSeen(t *testing.T) {
+	rest, flags := split([]string{"anchor", "umami", "database", "--from", "-"})
+	assert(t, len(rest) == 3, "the three positionals were not kept: %v", rest)
+	assert(t, len(flags) == 2, "the flag was not separated: %v", flags)
+
+	// And with no flag at all, which is the interactive form.
+	rest, flags = split([]string{"anchor", "umami", "database"})
+	assert(t, len(rest) == 3 && len(flags) == 0, "%v / %v", rest, flags)
+
+	// A flag before the positionals still works, because somebody will write it that way.
+	rest, flags = split([]string{"--from", "/tmp/x"})
+	assert(t, len(rest) == 0 && len(flags) == 2, "%v / %v", rest, flags)
+}
+
+// And the wiring, not just the helper.
+//
+// Testing `split` alone left the command able to ignore it entirely — removing the call changed
+// no test. This reaches secretCommand: with a `--from` naming a file that is not there, the
+// complaint must be about the file. A complaint about usage would mean the flag was never seen.
+func TestTheCommandItselfSeesTheFlag(t *testing.T) {
+	err := secretCommand(t.Context(),
+		[]string{"accept", "anchor", "umami", "database", "--from", "/nonexistent/nowhere"})
+	if err == nil {
+		t.Fatal("a missing file was accepted")
+	}
+	if strings.Contains(err.Error(), "secret accept <node>") {
+		t.Fatalf("the command did not see its flag and complained about usage instead: %v", err)
+	}
+}
+
+func assert(t *testing.T, ok bool, format string, args ...any) {
+	t.Helper()
+	if !ok {
+		t.Fatalf(format, args...)
+	}
+}

cmd/mesh-controller/sendable.go

@@ -0,0 +1,130 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"sort"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// sendable is a declaration as a machine is sent it: its resources and, for an adopted node, its
+// mode and which modules were taken on it (novox/hq ADR 0100).
+//
+// **Body is the only place the envelope is marshalled.** It was written by hand at every send site,
+// in the digest the mesh compares, and in `plan --json`; a key added at one and not another would
+// make a machine look out of date for ever, or send something `plan` never showed.
+type sendable struct {
+	Resources []map[string]any
+	// Sequence orders this send against every other to the same node: one higher each time, taken
+	// under the node's hold just before the body is made (novox/hq 04-ISSUES/107). Zero is not sent
+	// at all, which a host reads as "no order claimed" — the shape of every declaration before this.
+	Sequence int64
+	// Adoption is nil for a converged node, and then the body is byte for byte what it was before
+	// adoption existed: an older host parses the envelope strictly and would refuse the key.
+	Adoption *adoptionEnvelope
+
+	// Received and Mesh are not sent in the declaration. They are what this machine's memberships
+	// are issued with on the bus (novox/hq ADR 0167): each module's received contributions, from
+	// the same composition as its received files, and every machine's private-network address.
+	Received map[string]map[string][]catalogue.Contribution
+	Mesh     []string
+	// BusUsers is the bus's user list this declaration carries, empty for every machine but the one
+	// holding the bus; not sent apart from the file it is in. Its digest is recorded once sent, so
+	// whether that machine must go first is read from the list alone (novox/hq issue 249).
+	BusUsers string
+	// LeftOut is every module of the machine's set left out of this declaration because a stored
+	// setting cannot compose with its definition (novox/hq ADR 0163, rule 6), sorted. The host
+	// keeps that module's held things and touches none of its containers; a machine is told
+	// everything or nothing about what it IS told, and what it is not told is said. Absent from
+	// the body when empty, so a declaration that leaves nothing out is byte for byte what it was.
+	LeftOut []string
+	// leftOutWhy is why each was, for push and plan to say; never on the wire.
+	leftOutWhy map[string]string
+	// Builds is the build of each module this declaration carries — module to the commit its build
+	// was made from — recorded with the send and never on the wire (novox/hq issue 259, ADR 0221).
+	// Composed only on the send path; nil records that it is not known.
+	Builds map[string]string
+}
+
+// adoptionEnvelope is what an adopted node is told about its mode. Taken is every module taken on
+// it that it runs; Untaken is, for every module it runs that is not taken, the ids of every one of
+// that module's resources — what the host keeps as found until the module is taken (ADR 0103). Ids
+// rather than a rule to split them by, because a module's name may contain a dot.
+type adoptionEnvelope struct {
+	Taken   []string            `json:"taken"`
+	Untaken map[string][]string `json:"untaken,omitempty"`
+}
+
+// Body is the declaration's bytes, as sent and as digested.
+func (s sendable) Body() ([]byte, error) {
+	envelope := map[string]any{"declaration": 1, "resources": s.Resources}
+	if s.Adoption != nil {
+		envelope["adoption"] = s.Adoption
+	}
+	if s.Sequence > 0 {
+		envelope["sequence"] = s.Sequence
+	}
+	if len(s.LeftOut) > 0 {
+		envelope["left_out"] = s.LeftOut
+	}
+	// An empty declaration is deliberate here — the node owns nothing the mesh put there
+	// (novox/hq issue 127) — and the host refuses an empty body unless it is told the emptiness
+	// is meant, so a truncated or mis-composed body is never mistaken for "own nothing".
+	if len(s.Resources) == 0 {
+		envelope["owns_nothing"] = true
+	}
+	return json.Marshal(envelope)
+}
+
+// adoptionOf is the envelope for a node, nil when it is converged.
+func adoptionOf(ctx context.Context, inv *inventory.Inventory, record inventory.Node,
+	plan catalogue.Resolution, composed catalogue.Composed) (*adoptionEnvelope, error) {
+	if !record.Adopted {
+		return nil, nil
+	}
+	taken, err := inv.Taken(ctx, record.Name)
+	if err != nil {
+		return nil, err
+	}
+	return adoptionFor(plan, taken, composed), nil
+}
+
+// adoptionFor is the envelope computed from what was taken and who owns each resource.
+//
+// Every module the node runs that is not taken is untaken — including one pulled in by another
+// rather than assigned: what is found is kept until its module is taken, whoever put it there.
+func adoptionFor(plan catalogue.Resolution, taken []string,
+	composed catalogue.Composed) *adoptionEnvelope {
+	isTaken := map[string]bool{}
+	for _, m := range taken {
+		isTaken[m] = true
+	}
+	out := &adoptionEnvelope{Taken: []string{}}
+	runs := map[string]bool{}
+	for _, m := range plan.Modules {
+		runs[m.Module] = true
+		if isTaken[m.Module] {
+			out.Taken = append(out.Taken, m.Module)
+		}
+	}
+	sort.Strings(out.Taken)
+	for _, r := range composed.Resources {
+		id, _ := r["id"].(string)
+		module, owned := composed.Owner[id]
+		// Every kind, not only files and containers (novox/hq ADR 0103): a directory, a service, a
+		// container mounting what was found and an action run in a held container all reach what
+		// the machine already has. What the mesh declares of its own is never held.
+		if !owned || !runs[module] || isTaken[module] ||
+			strings.HasPrefix(id, catalogue.AdoptionPrefix) {
+			continue
+		}
+		if out.Untaken == nil {
+			out.Untaken = map[string][]string{}
+		}
+		out.Untaken[module] = append(out.Untaken[module], id)
+	}
+	return out
+}

cmd/mesh-controller/sendable_test.go

@@ -0,0 +1,443 @@
+package main
+
+import (
+	"bytes"
+	"encoding/json"
+	"io"
+	"os"
+	"reflect"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/overlay"
+)
+
+// novox/hq ADR 0100: every declaration an adopted node is sent says it is adopted and which modules
+// were taken on it; a converged node's declaration is byte for byte what it was.
+
+// helloWeb is a module the predecessor also runs: a page and a server under names it uses.
+func helloWeb() catalogue.Manifest {
+	return catalogue.Manifest{Module: "hello-web", Version: "1",
+		Resources: []map[string]any{
+			{"id": "page", "type": "file", "path": "/var/lib/hello-web/index.html", "content": "hello"},
+			{"id": "server", "type": "container", "name": "hello-web",
+				"image": "registry.example/hello@sha256:" + strings.Repeat("a", 64)},
+			{"id": "served", "type": "directory", "path": "/var/lib/hello-web"},
+		}}
+}
+
+// composed is what node would be sent now, as push composes it.
+func composed(t *testing.T, open *stores, node string) sendable {
+	t.Helper()
+	plan, settings, err := planFor(t.Context(), open, node)
+	if err != nil {
+		t.Fatal(err)
+	}
+	declared, err := declarationFor(t.Context(), open, node, plan, settings)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return declared
+}
+
+// convergedBefore is the envelope a converged node was sent before adoption existed, captured by
+// running this same composition at the commit this branch left main (0a39b7d). The mesh here is
+// aMesh's laptop with the private network taken off it, so nothing in the declaration is random:
+// what changes this string is a change to what a converged machine is sent, which is the thing an
+// older host would refuse.
+//
+// Re-captured 2026-10-01 (novox/hq issue 177): c978aa7 took `hosts` off every container — a
+// machine's own resolver knows the mesh's names now — and left this string carrying it, so the
+// guard failed for a day and nothing ran it. A field an older host never sees is the one change
+// this guard permits; a field it would refuse is the one it exists to catch.
+const convergedBefore = `{"declaration":1,"resources":[{"content":"hello","id":"hello-web.page","path":"/var/lib/hello-web/index.html","type":"file"},{"id":"hello-web.server","image":"registry.example/hello@sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa","name":"hello-web","type":"container"},{"id":"hello-web.served","path":"/var/lib/hello-web","type":"directory"}]}`
+
+func TestAConvergedDeclarationIsByteForByteWhatItWas(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, helloWeb())
+	if _, err := unassign(ctx, open, "laptop", overlay.Name); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := assign(ctx, open, "laptop", "hello-web"); err != nil {
+		t.Fatal(err)
+	}
+	declared := composed(t, open, "laptop")
+	if declared.Adoption != nil {
+		t.Fatal("a converged node was given an adoption envelope")
+	}
+	body, err := declared.Body()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(body) != convergedBefore {
+		t.Fatalf("a converged declaration changed; an older host parses this strictly:\n%s\n%s",
+			body, convergedBefore)
+	}
+	if bytes.Contains(body, []byte(`"adoption"`)) {
+		t.Fatal("a converged declaration names adoption; an older host would refuse it")
+	}
+}
+
+func TestAnAdoptedDeclarationCarriesItsModeAndWhatWasTaken(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, helloWeb())
+	if err := open.inventory.SetAdopted(ctx, "anchor", true); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := assign(ctx, open, "anchor", "hello-web"); err != nil {
+		t.Fatal(err)
+	}
+
+	declared := composed(t, open, "anchor")
+	if declared.Adoption == nil {
+		t.Fatal("an adopted node's declaration does not say it is adopted")
+	}
+	untaken := declared.Adoption.Untaken["hello-web"]
+	// Every kind — its directory too (novox/hq ADR 0103).
+	if !reflect.DeepEqual(untaken, []string{"hello-web.page", "hello-web.server",
+		"hello-web.served"}) {
+		t.Fatalf("hello-web's resources are not all named untaken: %v", declared.Adoption)
+	}
+	if len(declared.Adoption.Taken) != 0 {
+		t.Fatalf("nothing was taken, and the declaration says %v", declared.Adoption.Taken)
+	}
+
+	body, err := declared.Body()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var envelope map[string]any
+	if err := json.Unmarshal(body, &envelope); err != nil {
+		t.Fatal(err)
+	}
+	adoption, _ := envelope["adoption"].(map[string]any)
+	if _, ok := adoption["taken"].([]any); !ok {
+		t.Fatalf("taken is not a list on the wire, even empty: %s", body)
+	}
+
+	// The digest the mesh compares is the digest of what is sent: status and push agree.
+	would, err := wouldSend(ctx, open, mustNodes(t, open))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if would["anchor"] != digestOf(body) {
+		t.Fatal("the digest the mesh compares is not of the declaration push sends")
+	}
+
+	// plan --json prints that same envelope.
+	printed := stdoutOf(t, func() error { return planCommand(ctx, []string{"anchor", "--json"}) })
+	var compact bytes.Buffer
+	if err := json.Compact(&compact, []byte(printed)); err != nil {
+		t.Fatalf("plan --json is not JSON: %v\n%s", err, printed)
+	}
+	if digestOf(compact.Bytes()) != digestOf(body) {
+		t.Fatalf("plan --json shows something other than what push sends:\n%s", printed)
+	}
+
+	// Taking the module moves its resources out of untaken.
+	if err := open.inventory.Take(ctx, "anchor", "hello-web"); err != nil {
+		t.Fatal(err)
+	}
+	declared = composed(t, open, "anchor")
+	if _, still := declared.Adoption.Untaken["hello-web"]; still {
+		t.Fatalf("a taken module is still untaken: %v", declared.Adoption)
+	}
+	if !reflect.DeepEqual(declared.Adoption.Taken, []string{"hello-web"}) {
+		t.Fatalf("taken is %v", declared.Adoption.Taken)
+	}
+}
+
+func mustNodes(t *testing.T, open *stores) []inventory.Node {
+	t.Helper()
+	nodes, err := open.inventory.Nodes(t.Context())
+	if err != nil {
+		t.Fatal(err)
+	}
+	return nodes
+}
+
+// stdoutOf is what run printed.
+func stdoutOf(t *testing.T, run func() error) string {
+	t.Helper()
+	r, w, err := os.Pipe()
+	if err != nil {
+		t.Fatal(err)
+	}
+	saved := os.Stdout
+	os.Stdout = w
+	done := make(chan string)
+	go func() {
+		all, _ := io.ReadAll(r)
+		done <- string(all)
+	}()
+	runErr := run()
+	os.Stdout = saved
+	w.Close()
+	out := <-done
+	if runErr != nil {
+		t.Fatalf("%v\n%s", runErr, out)
+	}
+	return out
+}
+
+// novox/hq ADR 0100: a port a node was given for the store is where its consumers on other
+// machines are told to reach it, and a port given for the whole mesh is refused.
+func TestConsumersAreToldTheGivenPort(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, catalogue.Manifest{Module: "store", Version: "1",
+		Provides: []catalogue.Offer{{Name: "database", Scope: catalogue.ScopeMesh}},
+		Listens:  []catalogue.Listening{{Port: 5432, From: catalogue.FromMesh}},
+		Guards:   []int{5432},
+		Resources: []map[string]any{{"id": "server", "type": "container", "name": "mesh-store",
+			"ports": []any{"5432:5432"},
+			"image": "registry.example/pg@sha256:" + strings.Repeat("b", 64)}}})
+	register(t, open, catalogue.Manifest{Module: "app", Version: "1", Requires: []string{"database"}})
+	if _, err := assign(ctx, open, "anchor", "store"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := assign(ctx, open, "laptop", "app"); err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.SetSettings(ctx, "anchor", "store",
+		map[string]any{catalogue.PortsSetting: map[string]any{"5432": 5433}}); err != nil {
+		t.Fatal(err)
+	}
+
+	plan, _, err := planFor(ctx, open, "laptop")
+	if err != nil {
+		t.Fatal(err)
+	}
+	var told any
+	for _, n := range plan.Needs {
+		if n.Name == "database" {
+			told = n.Serves["port"]
+		}
+	}
+	if told != 5433 {
+		t.Fatalf("the consumer is told the store is on %v", told)
+	}
+	for _, r := range composed(t, open, "anchor").Resources {
+		if r["id"] == "store.server" && !reflect.DeepEqual(r["ports"], []any{"5433:5432"}) {
+			t.Fatalf("the store publishes %v", r["ports"])
+		}
+	}
+
+	// A port for the whole mesh is refused where it is set, not stored to refuse every node's
+	// declaration afterwards.
+	if err := open.inventory.SetSettings(ctx, "", "store",
+		map[string]any{catalogue.PortsSetting: map[string]any{"5432": 5434}}); err == nil ||
+		!strings.Contains(err.Error(), "per node") {
+		t.Fatalf("a port given for the whole mesh was not refused: %v", err)
+	}
+	plan, settings, err := planFor(ctx, open, "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := declarationFor(ctx, open, "anchor", plan, settings); err != nil {
+		t.Fatalf("the refused mesh-wide layer was stored anyway: %v", err)
+	}
+}
+
+// novox/hq ADR 0103: the guard an adopted node is sent follows what was taken there. A store
+// assigned but not taken is not guarded — its port may still be the predecessor's — and taking it
+// guards it from the next declaration.
+func TestTheGuardIsSentForTakenModulesOnly(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, catalogue.Manifest{Module: "store", Version: "1",
+		Listens: []catalogue.Listening{{Port: 5432, From: catalogue.FromMesh}},
+		Guards:  []int{5432},
+		Resources: []map[string]any{{"id": "server", "type": "container", "name": "mesh-store",
+			"ports": []any{"5432:5432"},
+			"image": "registry.example/pg@sha256:" + strings.Repeat("b", 64)}}})
+	if err := open.inventory.SetAdopted(ctx, "anchor", true); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := assign(ctx, open, "anchor", "store"); err != nil {
+		t.Fatal(err)
+	}
+	if hasID(composed(t, open, "anchor").Resources, catalogue.GuardID()) {
+		t.Fatal("an untaken store is guarded")
+	}
+	if err := open.inventory.Take(ctx, "anchor", "store"); err != nil {
+		t.Fatal(err)
+	}
+	for _, r := range composed(t, open, "anchor").Resources {
+		if r["id"] == catalogue.GuardID() {
+			if r["content"] != catalogue.AsGuard([]int{5432}) {
+				t.Fatalf("the taken store's guard is:\n%s", r["content"])
+			}
+			return
+		}
+	}
+	t.Fatal("a taken store is not guarded")
+}
+
+// novox/hq ADR 0038 and 0100: a module may publish a port the long way — `2222:22`, because the
+// machine's own ssh daemon holds 22 — and say it listens on the machine side of that mapping,
+// which is the number anything reaching it dials. A node moves that port by naming it, and the
+// number has to reach everything derived from it at once: what the runtime is handed, what an
+// adopted node is told to open, what its guard refuses, and what a consumer elsewhere dials.
+//
+// It reached none of them. The setting was refused outright for naming the machine side — so a
+// module's port could not be put back where the machine it replaces had it, and, worse, the
+// node's every push failed for as long as the setting existed.
+func TestTheMachineSideOfAMappingIsMovedEverywhereTheNumberIsUsed(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, catalogue.Manifest{Module: "forge", Version: "1",
+		Provides: []catalogue.Offer{{Name: "git-over-ssh", Scope: catalogue.ScopeMesh}},
+		Serves:   map[string]map[string]any{"git-over-ssh": {"port": 2222}},
+		Listens:  []catalogue.Listening{{Port: 2222, From: catalogue.FromMesh}},
+		Resources: []map[string]any{{"id": "server", "type": "container", "name": "forge",
+			"ports": []any{"2222:22"},
+			"image": "registry.example/forge@sha256:" + strings.Repeat("c", 64)}}})
+	register(t, open, catalogue.Manifest{Module: "app", Version: "1",
+		Requires: []string{"git-over-ssh"}})
+	if err := open.inventory.SetAdopted(ctx, "anchor", true); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := assign(ctx, open, "anchor", "forge"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := assign(ctx, open, "laptop", "app"); err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.Take(ctx, "anchor", "forge"); err != nil {
+		t.Fatal(err)
+	}
+	if err := open.inventory.SetSettings(ctx, "anchor", "forge",
+		map[string]any{catalogue.PortsSetting: map[string]any{"2222": 222}}); err != nil {
+		t.Fatal(err)
+	}
+
+	resources := composed(t, open, "anchor").Resources
+	var container, opening map[string]any
+	for _, r := range resources {
+		switch r["id"] {
+		case "forge.server":
+			container = r
+		case catalogue.OpeningID("tcp", 222, catalogue.PathForwarded):
+			opening = r
+		}
+		if id, _ := r["id"].(string); strings.HasPrefix(id, "adoption.opening-tcp-2222-") {
+			t.Errorf("the adopted node is told to open the port the forge was moved off: %s", id)
+		}
+	}
+	if container == nil || !reflect.DeepEqual(container["ports"], []any{"222:22"}) {
+		t.Fatalf("the forge's container publishes %v", container["ports"])
+	}
+	if opening == nil || opening["to"] != 22 || opening["from"] != catalogue.OpeningFromMesh {
+		t.Fatalf("no opening for the port this node gave the forge: %v", opening)
+	}
+	var guard map[string]any
+	for _, r := range resources {
+		if r["id"] == catalogue.GuardID() {
+			guard = r
+		}
+	}
+	if guard == nil || guard["content"] != catalogue.AsGuard([]int{222}) {
+		t.Fatalf("the guard does not refuse the port the forge is on:\n%v", guard["content"])
+	}
+
+	// And the consumer on the other machine dials the same number.
+	plan, _, err := planFor(ctx, open, "laptop")
+	if err != nil {
+		t.Fatal(err)
+	}
+	var told any
+	for _, n := range plan.Needs {
+		if n.Name == "git-over-ssh" {
+			told = n.Serves["port"]
+		}
+	}
+	if told != 222 {
+		t.Fatalf("the consumer is told the forge answers on %v", told)
+	}
+}
+
+func TestAnEmptyDeclarationSaysOwnsNothing(t *testing.T) {
+	// The host refuses an empty body unless told the emptiness is meant (novox/hq issue 127).
+	body, err := sendable{}.Body()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var env map[string]any
+	if err := json.Unmarshal(body, &env); err != nil {
+		t.Fatal(err)
+	}
+	if env["owns_nothing"] != true {
+		t.Fatalf("an empty declaration must mark owns_nothing; got %v", env)
+	}
+	// A declaration with resources does not carry the marker.
+	body, _ = sendable{Resources: []map[string]any{{"id": "x"}}}.Body()
+	var env2 map[string]any
+	_ = json.Unmarshal(body, &env2)
+	if _, present := env2["owns_nothing"]; present {
+		t.Fatalf("a non-empty declaration must not mark owns_nothing; got %v", env)
+	}
+}
+
+// A setting is judged where it is stored, and an impossible one costs a module, not a machine
+// (novox/hq ADR 0163, rule 6): stored while it composed, a setting whose definition then moved from
+// under it leaves that module out of the declaration — said in the envelope, so the host keeps the
+// module's things — and the machine is told everything else.
+func TestADefinitionMovingUnderAStoredSettingLeavesThatModuleOutNotTheMachine(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	web := helloWeb()
+	web.Resources[1]["ports"] = []any{"8080"}
+	register(t, open, web)
+	register(t, open, catalogue.Manifest{Module: "notes", Version: "1",
+		Resources: []map[string]any{{"id": "conf", "type": "file", "path": "/etc/notes.conf", "content": "x"}}})
+	for _, m := range []string{"hello-web", "notes"} {
+		if _, err := assign(ctx, open, "laptop", m); err != nil {
+			t.Fatal(err)
+		}
+	}
+	// Judged where it is stored: a port the module does not publish is refused by name.
+	err := open.inventory.SetSettings(ctx, "laptop", "hello-web",
+		map[string]any{catalogue.PortsSetting: map[string]any{"9999": 10000}})
+	if err == nil || !strings.Contains(err.Error(), "hello-web on laptop") || !strings.Contains(err.Error(), "9999") {
+		t.Fatalf("an impossible setting was stored: %v", err)
+	}
+	if err := open.inventory.SetSettings(ctx, "laptop", "hello-web",
+		map[string]any{catalogue.PortsSetting: map[string]any{"8080": 10000}}); err != nil {
+		t.Fatal(err)
+	}
+	if declared := composed(t, open, "laptop"); len(declared.LeftOut) != 0 {
+		t.Fatalf("a setting that composes left a module out: %v", declared.LeftOut)
+	}
+
+	// The definition moves: the container publishes another port now.
+	web.Version = "2"
+	web.Resources[1]["ports"] = []any{"9090"}
+	register(t, open, web)
+	declared := composed(t, open, "laptop")
+	if len(declared.LeftOut) != 1 || declared.LeftOut[0] != "hello-web" {
+		t.Fatalf("hello-web is not left out: %v", declared.LeftOut)
+	}
+	if !strings.Contains(declared.leftOutWhy["hello-web"], "no container of its publishes 8080") {
+		t.Fatalf("why it was left out is not said: %v", declared.leftOutWhy)
+	}
+	if hasID(declared.Resources, "hello-web.server") || !hasID(declared.Resources, "notes.conf") {
+		t.Fatalf("the machine was not told everything else: %v", declared.Resources)
+	}
+	body, err := declared.Body()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var env map[string]any
+	if err := json.Unmarshal(body, &env); err != nil {
+		t.Fatal(err)
+	}
+	left, _ := env["left_out"].([]any)
+	if len(left) != 1 || left[0] != "hello-web" {
+		t.Fatalf("the envelope does not say what was left out: %v", env)
+	}
+}

cmd/mesh-controller/sequence_test.go

@@ -0,0 +1,77 @@
+package main
+
+import (
+	"encoding/json"
+	"testing"
+)
+
+// A declaration's only identity was the digest of its bytes; the controller already held a per-node
+// lock and recorded each send, so the order existed and was thrown away at the wire (novox/hq
+// 04-ISSUES/107).
+
+func TestASendCarriesItsNumberInsideTheSignedBytes(t *testing.T) {
+	body, err := sendable{Resources: []map[string]any{{"id": "x", "type": "file"}}, Sequence: 7}.Body()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var env map[string]any
+	if err := json.Unmarshal(body, &env); err != nil {
+		t.Fatal(err)
+	}
+	if got, _ := env["sequence"].(float64); got != 7 {
+		t.Fatalf("the body carries sequence %v, wanted 7", env["sequence"])
+	}
+}
+
+func TestAnUnnumberedSendIsByteForByteWhatItWasBefore(t *testing.T) {
+	// Zero is not sent at all. A host reads absence as "no order claimed" — the shape of every
+	// declaration before this — so an older host, or the read-only comparison against a machine
+	// sent nothing since sends were numbered, sees exactly the bytes it always saw.
+	body, err := sendable{Resources: []map[string]any{{"id": "x", "type": "file"}}}.Body()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var env map[string]any
+	if err := json.Unmarshal(body, &env); err != nil {
+		t.Fatal(err)
+	}
+	if _, present := env["sequence"]; present {
+		t.Fatalf("a send numbered zero put a sequence on the wire: %s", body)
+	}
+}
+
+func TestEachSendToANodeIsOneHigherAndReadable(t *testing.T) {
+	open := aMesh(t)
+	record, err := open.inventory.NodeByName(t.Context(), "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	// Sent nothing since numbering existed: what it would be sent is composed with zero, which is
+	// not on the wire, which is what it was actually sent.
+	if n, err := open.inventory.Sequence(t.Context(), record.ID); err != nil || n != 0 {
+		t.Fatalf("a fresh node reads sequence %d, %v", n, err)
+	}
+	first, err := open.inventory.NextSequence(t.Context(), record.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	second, err := open.inventory.NextSequence(t.Context(), record.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if first != 1 || second != 2 {
+		t.Fatalf("two sends were numbered %d and %d", first, second)
+	}
+	// And the read path sees the last one taken, so the comparison composes what was sent.
+	if n, err := open.inventory.Sequence(t.Context(), record.ID); err != nil || n != 2 {
+		t.Fatalf("after two sends the node reads sequence %d, %v", n, err)
+	}
+	// Another node counts on its own.
+	other, err := open.inventory.NodeByName(t.Context(), "laptop")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if n, err := open.inventory.NextSequence(t.Context(), other.ID); err != nil || n != 1 {
+		t.Fatalf("a second node's first send was numbered %d, %v", n, err)
+	}
+}

cmd/mesh-controller/source.go

@@ -0,0 +1,171 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"strconv"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// where a build's repository is (novox/hq ADR 0111).
+//
+// A repository is on the mesh's own forge, or it is anywhere else. The first is recorded as its path
+// on the forge holding the git seat, and cloned from wherever that forge runs at the moment of
+// building; the second is a URL, recorded and cloned exactly as given. The build machine is not told
+// the difference — it is handed a URL either way — because only the control plane knows where the
+// seat's holder runs.
+
+// gitSeat is the seat a self-hosted repository lives on.
+const gitSeat = "git"
+
+// buildSource is where a build's repository is: a URL, or a path on a seat's holder.
+type buildSource struct {
+	Repository string
+	Seat       string
+}
+
+// String is the source as a person reads it, which for one on a seat is not the URL: the URL is a
+// fact about where the forge happens to run today.
+func (s buildSource) String() string {
+	if s.Seat == "" {
+		return s.Repository
+	}
+	return fmt.Sprintf("%s on the %s seat", s.Repository, s.Seat)
+}
+
+// onASeat refuses an address given as a path on the forge.
+//
+// **A URL here would be recorded as a path**, and then composed onto the forge's address as one —
+// cloning `http://forge:3000/https://github.com/…`. Refused by what an address plainly looks like,
+// not repaired: `--self` promises a path, and something that is not one is a mistake to name.
+func onASeat(repository string) error {
+	if strings.Contains(repository, ":") || strings.HasPrefix(repository, "/") ||
+		strings.Trim(repository, "/") == "" {
+		return fmt.Errorf("--self takes the repository's path on the forge, such as novox/mesh-catalog, "+
+			"and %q is not one — without --self it is built from exactly what is given", repository)
+	}
+	return nil
+}
+
+// cloneFrom is the URL a build machine clones for a source.
+//
+// A URL is itself. A path on a seat is composed from the seat's holder as the mesh sees it now —
+// the same view planning takes of every machine, so the forge a build clones from is the forge the
+// mesh says holds the seat.
+func cloneFrom(ctx context.Context, source buildSource) (string, error) {
+	if source.Seat == "" {
+		return source.Repository, nil
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return "", err
+	}
+	defer open.Close()
+	shelf, err := open.inventory.Catalogue(ctx)
+	if err != nil {
+		return "", err
+	}
+	world, err := theRestOfTheMesh(ctx, open.inventory, shelf, "")
+	if err != nil {
+		return "", err
+	}
+	return clonedFromSeat(world, source.Seat, source.Repository)
+}
+
+// clonedFromSeat composes the clone URL for a repository on a seat's holder.
+//
+// **Refused, never defaulted, at every step that has no answer.** Nobody holding the seat is a mesh
+// without a forge of its own: it builds from external repositories and must say so rather than fail
+// to clone. A holder off the private network cannot be reached by any build machine. A holder that
+// serves no scheme or port has nothing to compose from — a default port here would be the forge's
+// address guessed, which is the thing this exists to stop.
+func clonedFromSeat(world catalogue.World, seatName, repository string) (string, error) {
+	base, err := seatBase(world, seatName)
+	if err != nil {
+		return "", err
+	}
+	path := strings.TrimSuffix(strings.Trim(repository, "/"), ".git")
+	return fmt.Sprintf("%s/%s.git", base, path), nil
+}
+
+// seatBase is `scheme://host:port` of a seat's holder as the mesh reaches it, for cloning.
+func seatBase(world catalogue.World, seatName string) (string, error) {
+	seat, known := catalogue.SeatNamed(seatName)
+	if !known || seat.Delivers == "" {
+		return "", fmt.Errorf("%q is not a seat a repository can live on", seatName)
+	}
+	var holder *catalogue.Held
+	for i, h := range world.Held {
+		if h.Claim == seat.Name && h.Scope == seat.Scope {
+			holder = &world.Held[i]
+			break
+		}
+	}
+	if holder == nil {
+		return "", fmt.Errorf("nobody holds the %s seat, so nothing can be cloned from this mesh's "+
+			"forge — assign a module that claims it, or build from the repository's URL without --self",
+			seat.Name)
+	}
+	var provider *catalogue.Provider
+	for i, p := range world.Offered[seat.Delivers] {
+		if p.Node == holder.Node && p.Module == holder.Module {
+			provider = &world.Offered[seat.Delivers][i]
+		}
+	}
+	if provider == nil {
+		return "", fmt.Errorf("%s on %s holds the %s seat and offers no %q to clone from",
+			holder.Module, holder.Node, seat.Name, seat.Delivers)
+	}
+	if provider.At == "" {
+		return "", fmt.Errorf("%s on %s holds the %s seat and is not on the private network, so no "+
+			"build machine can reach it", holder.Module, holder.Node, seat.Name)
+	}
+	scheme, _ := provider.Serves["scheme"].(string)
+	port := servedPort(provider.Serves["port"])
+	if scheme == "" || port == "" {
+		return "", fmt.Errorf("%s on %s holds the %s seat and does not serve a scheme and a port for %q",
+			holder.Module, holder.Node, seat.Name, seat.Delivers)
+	}
+	return fmt.Sprintf("%s://%s:%s", scheme, provider.At, port), nil
+}
+
+// seatBases is the clone base of every seat a recipe's context may name, for a build request
+// (novox/hq ADR 0155). A seat nobody holds is left out rather than refused here: the build may not
+// name it at all, and if it does the builder refuses with the seat's name.
+func seatBases(ctx context.Context) map[string]string {
+	open, err := openStores(ctx)
+	if err != nil {
+		return nil
+	}
+	defer open.Close()
+	shelf, err := open.inventory.Catalogue(ctx)
+	if err != nil {
+		return nil
+	}
+	world, err := theRestOfTheMesh(ctx, open.inventory, shelf, "")
+	if err != nil {
+		return nil
+	}
+	bases := map[string]string{}
+	for _, seatName := range []string{gitSeat} {
+		if base, err := seatBase(world, seatName); err == nil {
+			bases[seatName] = base
+		}
+	}
+	return bases
+}
+
+// servedPort is a served port as text, however the manifest and the node's settings carried it.
+func servedPort(v any) string {
+	switch p := v.(type) {
+	case float64:
+		return strconv.Itoa(int(p))
+	case int:
+		return strconv.Itoa(p)
+	case string:
+		return p
+	}
+	return ""
+}

cmd/mesh-controller/source_test.go

@@ -0,0 +1,108 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// Defends novox/hq ADR 0111: a build source is on the git seat, or it is external.
+
+func forgeHolding(port any) catalogue.World {
+	return catalogue.World{
+		Held: []catalogue.Held{{Claim: "git", Scope: catalogue.ScopeMesh, Node: "anchor", Module: "gitea"}},
+		Offered: map[string][]catalogue.Provider{"git": {
+			// A second forge that does not hold the seat, so taking the first one found would be wrong.
+			{Node: "archive", At: "archive.internal", Module: "gitea-mirror",
+				Serves: map[string]any{"scheme": "http", "port": float64(3000)}},
+			{Node: "anchor", At: "anchor.internal", Module: "gitea",
+				Serves: map[string]any{"scheme": "http", "port": port}},
+		}},
+	}
+}
+
+func TestARepositoryOnTheSeatIsClonedFromItsHolder(t *testing.T) {
+	got, err := clonedFromSeat(forgeHolding(float64(3000)), "git", "novox/mesh-catalog")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got != "http://anchor.internal:3000/novox/mesh-catalog.git" {
+		t.Fatalf("cloned from %s", got)
+	}
+}
+
+func TestAMovedForgeIsFollowedWithoutRewritingAnything(t *testing.T) {
+	// The whole point: the node gave the forge another port, and the same recorded path clones
+	// from the new one. Nothing recorded contained the old one to be wrong.
+	got, err := clonedFromSeat(forgeHolding(float64(3100)), "git", "novox/mesh-catalog")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(got, ":3100/") {
+		t.Fatalf("the moved port was not followed: %s", got)
+	}
+}
+
+func TestWithNobodyHoldingTheSeatASelfHostedBuildIsRefusedAndSaysWhy(t *testing.T) {
+	_, err := clonedFromSeat(catalogue.World{}, "git", "novox/mesh-catalog")
+	if err == nil {
+		t.Fatal("a repository was cloned from a forge the mesh does not have")
+	}
+	for _, want := range []string{"nobody holds the git seat", "without --self"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Fatalf("the refusal does not say %q: %v", want, err)
+		}
+	}
+}
+
+func TestAnExternalRepositoryIsClonedExactlyAsGiven(t *testing.T) {
+	// Unaffected by the seat, held or not: GitHub and GitLab are the ordinary cases.
+	given := "https://github.com/someone/something.git"
+	got, err := cloneFrom(t.Context(), buildSource{Repository: given})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got != given {
+		t.Fatalf("an external repository became %s", got)
+	}
+}
+
+func TestAHolderOffThePrivateNetworkIsRefused(t *testing.T) {
+	world := forgeHolding(float64(3000))
+	world.Offered["git"][1].At = ""
+	if _, err := clonedFromSeat(world, "git", "novox/mesh-catalog"); err == nil ||
+		!strings.Contains(err.Error(), "private network") {
+		t.Fatalf("a forge nothing can reach was cloned from: %v", err)
+	}
+}
+
+func TestAHolderServingNoPortIsRefusedRatherThanGuessed(t *testing.T) {
+	// A default port would be the forge's address guessed, which is what this exists to stop.
+	if _, err := clonedFromSeat(forgeHolding(nil), "git", "novox/mesh-catalog"); err == nil {
+		t.Fatal("a port was guessed for a forge that serves none")
+	}
+}
+
+func TestAnAddressGivenAsAPathOnTheForgeIsRefused(t *testing.T) {
+	for _, bad := range []string{
+		"https://github.com/someone/something.git",
+		"git@anchor:novox/mesh-catalog.git",
+		"/srv/git/mesh-catalog",
+		"",
+	} {
+		if err := onASeat(bad); err == nil {
+			t.Errorf("--self accepted %q as a path on the forge", bad)
+		}
+	}
+	if err := onASeat("novox/mesh-catalog"); err != nil {
+		t.Errorf("a path on the forge was refused: %v", err)
+	}
+}
+
+func TestASourceOnTheSeatReadsAsAPathNotAnAddress(t *testing.T) {
+	s := buildSource{Repository: "novox/mesh-catalog", Seat: "git"}
+	if got := s.String(); got != "novox/mesh-catalog on the git seat" {
+		t.Fatalf("read as %q", got)
+	}
+}

cmd/mesh-controller/standing.go

@@ -0,0 +1,120 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A provider that keeps failing a consumer is a problem the controller reports (novox/hq ADR 0224).
+//
+// On 2026-10-05 the identity provider's provisioner failed every consumer from shortly after midnight
+// until it was fixed by hand that night — 31,000 refused logins after its database was moved and its
+// admin kept an older password — and `status` called the mesh well all day (novox/hq issue 179). A
+// provider now announces a consumer it has failed for minutes; the controller keeps it until the
+// provider says it recovered; and `status`, its JSON and `node show` name it, breaking "all well".
+
+// standings keeps what providers say, in the inventory.
+type standings struct{ inv *inventory.Inventory }
+
+func (s standings) Stood(ctx context.Context, st link.Standing) (bool, error) {
+	return s.inv.KeepStanding(ctx, st.Failing, inventory.ProviderStanding{
+		Module: st.Module, ProviderNode: st.ProviderNode, Provision: st.Provider,
+		Consumer: st.Consumer, ConsumerNode: st.Node,
+		Class: st.Class, Error: st.Error, Since: st.Since, Attempts: st.Attempts,
+	})
+}
+
+// failingProviders is every consumer a provider still assigned where it ran says it keeps failing.
+//
+// **A provider no longer assigned is not asked about.** Its last word stays in the store, and is
+// not a problem: nothing runs there to fail anybody. Assigned again, its first success for each
+// consumer clears it.
+func failingProviders(ctx context.Context, inv *inventory.Inventory) ([]inventory.ProviderStanding, error) {
+	all, err := inv.FailingProviders(ctx)
+	if err != nil {
+		return nil, fmt.Errorf("what providers say they keep failing cannot be read: %w", err)
+	}
+	assigned := map[string]map[string]bool{}
+	var out []inventory.ProviderStanding
+	for _, s := range all {
+		on, asked := assigned[s.ProviderNode]
+		if !asked {
+			modules, err := inv.Assigned(ctx, s.ProviderNode)
+			if err != nil {
+				// A provider on a machine the mesh no longer knows has nothing running to fail anybody.
+				modules = nil
+			}
+			on = map[string]bool{}
+			for _, m := range modules {
+				on[m] = true
+			}
+			assigned[s.ProviderNode] = on
+		}
+		if on[s.Module] {
+			out = append(out, s)
+		}
+	}
+	return out, nil
+}
+
+// failingLines is how status says them: one consumer per entry, the error under it, and a provider
+// that stopped repeating itself said so.
+func failingLines(list []inventory.ProviderStanding, now time.Time) []string {
+	var out []string
+	for _, s := range list {
+		where := s.Module
+		if s.ProviderNode != "" {
+			where += " on " + s.ProviderNode
+		}
+		whom := s.Consumer
+		if s.ConsumerNode != "" {
+			whom += " (" + s.ConsumerNode + ")"
+		}
+		out = append(out, fmt.Sprintf("  %-24s fails %s: %s, for %s (%d attempts since %s)",
+			where, whom, orUnclassed(s.Class), roughly(now.Sub(s.Since)), s.Attempts,
+			s.Since.Local().Format("2006-01-02 15:04")))
+		if e := strings.TrimSpace(s.Error); e != "" {
+			out = append(out, fmt.Sprintf("  %-24s %s", "", firstLine(e)))
+		}
+		if s.Quiet(now) {
+			out = append(out, fmt.Sprintf("  %-24s not said again for %s — the provider has stopped "+
+				"saying anything, so this is its last word", "", roughly(now.Sub(s.SaidAt))))
+		}
+	}
+	return out
+}
+
+func orUnclassed(class string) string {
+	if class == "" {
+		return "failing"
+	}
+	return class
+}
+
+// printFailing is the status section, said when there is anything to say.
+func printFailing(list []inventory.ProviderStanding, now time.Time) {
+	if len(list) == 0 {
+		return
+	}
+	fmt.Printf("%d consumer(s) a provider keeps failing (ADR 0224):\n\n", len(list))
+	for _, line := range failingLines(list, now) {
+		fmt.Println(line)
+	}
+	fmt.Printf("\n  the provider's journal has every attempt; it says recovered on its next success\n\n")
+}
+
+// failingOn is the standings that concern one machine: a provider running there, or a consumer.
+func failingOn(list []inventory.ProviderStanding, node string) []inventory.ProviderStanding {
+	var out []inventory.ProviderStanding
+	for _, s := range list {
+		if s.ProviderNode == node || s.ConsumerNode == node {
+			out = append(out, s)
+		}
+	}
+	return out
+}

cmd/mesh-controller/standing_test.go

@@ -0,0 +1,115 @@
+package main
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A provider that keeps failing a consumer is a problem `status` names (novox/hq ADR 0224). On
+// 2026-10-05 the identity provider refused every consumer for a day and status called the mesh well
+// (04-ISSUES/179): this is that day, told to the controller the way the provider now tells it.
+func TestAProviderFailingAConsumerBreaksAllWellUntilItRecovers(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, catalogue.Manifest{Module: "idp", Version: "1",
+		Receives: map[string]string{"oidc-client": "/var/lib/mesh/idp/mesh.json"}})
+	if _, err := assign(ctx, open, "anchor", "idp"); err != nil {
+		t.Fatal(err)
+	}
+	kept := standings{open.inventory}
+	since := time.Now().Add(-23 * time.Hour)
+	failing := link.Standing{Module: "idp", Failing: true, Provider: "oidc-client", ProviderNode: "anchor",
+		Consumer: "mesh_laptop_dashboard", Node: "laptop", Class: "credentials-rejected",
+		Error: `Keycloak token request failed: 401 {"error":"invalid_grant"}`, Since: since, Attempts: 31000}
+	if _, err := kept.Stood(ctx, failing); err != nil {
+		t.Fatal(err)
+	}
+
+	asked, err := theThreeQuestions(ctx, open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if asked.well() {
+		t.Fatal("a mesh whose identity provider fails a consumer reads as well")
+	}
+	said := printed(t, func() error { return printStatus(asked) })
+	for _, want := range []string{"1 consumer(s) a provider keeps failing", "idp on anchor",
+		"mesh_laptop_dashboard (laptop)", "credentials-rejected", "31000 attempts", "invalid_grant"} {
+		if !strings.Contains(said, want) {
+			t.Fatalf("status does not say %q:\n%s", want, said)
+		}
+	}
+	if strings.Contains(said, "all doing what they were told") {
+		t.Fatalf("status said all well beside a failing provider:\n%s", said)
+	}
+	body, err := statusAsJSON(asked)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var doc struct {
+		Failing []inventory.ProviderStanding `json:"failing"`
+	}
+	if err := json.Unmarshal(body, &doc); err != nil || len(doc.Failing) != 1 || doc.Failing[0].Consumer != "mesh_laptop_dashboard" {
+		t.Fatalf("the document does not carry it: %v\n%s", err, body)
+	}
+	// Both machines' `node show` name it: where the provider runs, and where the consumer is.
+	for _, node := range []string{"anchor", "laptop"} {
+		shown := printed(t, func() error { return showNode(ctx, open.inventory, node) })
+		if !strings.Contains(shown, "a provider keeps failing") || !strings.Contains(shown, "mesh_laptop_dashboard") {
+			t.Fatalf("node show %s does not name it:\n%s", node, shown)
+		}
+	}
+
+	// Recovered: gone, and the mesh may be well again as far as this is concerned.
+	failing.Failing = false
+	if cleared, err := kept.Stood(ctx, failing); err != nil || !cleared {
+		t.Fatalf("%v %v", cleared, err)
+	}
+	asked, err = theThreeQuestions(ctx, open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(asked.failing) != 0 {
+		t.Fatalf("a recovered consumer is still named: %+v", asked.failing)
+	}
+}
+
+// A provider no longer assigned where it ran has nothing running to fail anybody: its last word is
+// not a problem.
+func TestAnUnassignedProvidersLastWordIsNotAProblem(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	if _, err := (standings{open.inventory}).Stood(ctx, link.Standing{Module: "gone", Failing: true,
+		ProviderNode: "anchor", Consumer: "x", Since: time.Now()}); err != nil {
+		t.Fatal(err)
+	}
+	asked, err := theThreeQuestions(ctx, open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(asked.failing) != 0 {
+		t.Fatalf("%+v", asked.failing)
+	}
+}
+
+func TestAProviderThatStoppedRepeatingItselfIsSaidToHaveGoneQuiet(t *testing.T) {
+	now := time.Now()
+	lines := strings.Join(failingLines([]inventory.ProviderStanding{{
+		Module: "idp", ProviderNode: "anchor", Consumer: "c", Class: "unreachable", Error: "connection refused\nmore",
+		Since: now.Add(-3 * time.Hour), SaidAt: now.Add(-2 * time.Hour), Attempts: 9,
+	}}, now), "\n")
+	for _, want := range []string{"unreachable, for 3h", "connection refused", "not said again for 2h"} {
+		if !strings.Contains(lines, want) {
+			t.Fatalf("%q not in:\n%s", want, lines)
+		}
+	}
+	if strings.Contains(lines, "more") {
+		t.Fatalf("more than the first line of an error:\n%s", lines)
+	}
+}

cmd/mesh-controller/status.go

@@ -0,0 +1,572 @@
+package main
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/overlay"
+)
+
+// is anything broken, is anything not answering, is anything out of date.
+//
+// Split out of main.go, which had reached 2,769 lines because appending was always the
+// cheapest next step. That is how novox/hq ADR 0001 records `hal/sdk` reaching 34,636:
+// nothing in it was wrong, and no one edit was the one that should have been a new file.
+
+// short is a commit as a person refers to it.
+func short(commit string) string {
+	if len(commit) > 8 {
+		return commit[:8]
+	}
+	return commit
+}
+
+// statusCommand answers "did my change go out?".
+//
+// novox/hq ADR 0010 names losing that question as the real risk of replacing a pipeline with a
+// comparison: it is answerable today by opening a pipeline, and something has to replace that or
+// this is worse to live with whatever its other properties.
+//
+// The answer is not "a job succeeded". It is which modules the mesh has not built from what their
+// source now has, and which machines are running the old one.
+func statusCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("status", flag.ContinueOnError)
+	asJSON := set.Bool("json", false, "the same answers, for something other than a person")
+	if _, err := parseAround(set, args); err != nil {
+		return err
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	return statusFor(ctx, open, *asJSON)
+}
+
+// statusFor asks and answers, against stores somebody else opened.
+//
+// Split from the command so what it prints can be read by a test. The sentence it prints when nothing
+// is wrong has been acted on and been misleading (novox/hq 04-ISSUES/145, 125), which makes its exact
+// words the thing worth holding still.
+func statusFor(ctx context.Context, open *stores, asJSON bool) error {
+	asked, err := theThreeQuestions(ctx, open)
+	if err != nil {
+		return err
+	}
+
+	if asJSON {
+		body, err := statusAsJSON(asked)
+		if err != nil {
+			return err
+		}
+		fmt.Println(string(body))
+		return nil
+	}
+	return printStatus(asked)
+}
+
+// printStatus is the words, separated from the questions.
+//
+// **Its exact sentences have been acted on and been misleading twice** — a held module reading as a
+// machine doing what it was told (novox/hq 04-ISSUES/125), and "all doing what they were told" being
+// true of a mesh in which no module could reach another (04-ISSUES/145). So they are written where a
+// test can read them without a store, a bus or a machine.
+func printStatus(asked answers) error {
+	wrong, nodes, quiet := asked.wrong, asked.nodes, asked.quiet
+	behind, sources := asked.behind, asked.sources
+
+	if len(asked.refused) > 0 {
+		// First, above everything else. A machine that cannot be worked out is not running an old
+		// declaration — it has no declaration, and nothing below this line is about it.
+		var names []string
+		for name := range asked.refused {
+			names = append(names, name)
+		}
+		sort.Strings(names)
+		fmt.Printf("%d machine(s) cannot be worked out at all, so nothing can be sent to them:\n\n",
+			len(names))
+		for _, name := range names {
+			fmt.Printf("  %s\n", name)
+			for _, line := range strings.Split(strings.TrimRight(asked.refused[name], "\n"), "\n") {
+				fmt.Printf("    %s\n", strings.TrimSpace(line))
+			}
+		}
+		fmt.Println()
+	}
+
+	if asked.network != "" {
+		fmt.Printf("the private network could not be computed:\n  %s\n\n",
+			strings.ReplaceAll(strings.TrimRight(asked.network, "\n"), "\n", "\n  "))
+	}
+
+	if len(wrong) > 0 {
+		fmt.Printf("%d machine(s) are not doing what they were told:\n\n", len(wrong))
+		for _, d := range wrong {
+			fmt.Printf("  %-18s %-9s %s\n", d.Node, d.Outcome, d.At.Local().Format("2006-01-02 15:04"))
+			if d.Refused != "" {
+				// The host's own words. It says exactly what it could not accept, and nothing
+				// written here would say it better.
+				fmt.Printf("  %-18s %s\n", "", firstLine(d.Refused))
+			}
+			for _, f := range d.Failed {
+				fmt.Printf("  %-18s %s: %s\n", "", f.ID, firstLine(f.Error))
+			}
+			if d.Stuck() {
+				// Said apart from the failure itself. The host's words say what is wrong; this
+				// says it is not new — the machine has applied, failed the same way and reported
+				// so this many times, and will keep doing exactly that until something changes
+				// (novox/hq 04-ISSUES/065).
+				fmt.Printf("  %-18s stuck: the same failure %d times since %s — it will not fix itself\n",
+					"", d.Times, d.Since.Local().Format("2006-01-02 15:04"))
+			}
+		}
+		fmt.Println()
+	}
+
+	// A provider failing a consumer, beside machines failing what they were told: both are something
+	// not working now (novox/hq ADR 0224).
+	printFailing(asked.failing, time.Now())
+
+	if len(quiet) > 0 {
+		var said []string
+		for _, n := range quiet {
+			said = append(said, n.Name+" ("+heardFrom(n)+")")
+		}
+		fmt.Printf("%d machine(s) not heard from lately:\n  %s\n\n",
+			len(quiet), strings.Join(said, "\n  "))
+	}
+
+	if open, late := openPlans(asked.plans, asked.paused); len(open) > 0 {
+		fmt.Printf("%d plan(s) open", len(open))
+		if late > 0 {
+			fmt.Printf(", %d waiting past %s", late, planWaitBound)
+		}
+		fmt.Println(":")
+		for _, p := range open {
+			fmt.Printf("  %s\n", planLineWith(p, time.Now(), asked.paused))
+		}
+		fmt.Println()
+	}
+
+	if len(behind) > 0 {
+		var names []string
+		for m := range behind {
+			names = append(names, m)
+		}
+		sort.Strings(names)
+
+		fmt.Printf("%d module(s) behind their source:\n\n", len(behind))
+		for _, m := range names {
+			from := sources[m]
+			fmt.Printf("  %-18s holds %s, source has %s\n", m, short(from.BuiltFrom), short(from.Head))
+			if on := behind[m]; len(on) > 0 {
+				// The part somebody actually wants. A module being out of date is a fact about
+				// the catalogue; machines running the old one is the thing with consequences.
+				fmt.Printf("  %-18s running on %s\n", "", strings.Join(on, ", "))
+			} else {
+				fmt.Printf("  %-18s assigned to nothing\n", "")
+			}
+		}
+		// The remedy, beside the problem. A status that says what is wrong and not what to do
+		// about it makes somebody go and find the command, and the command is the whole point of
+		// having noticed.
+		fmt.Printf("\n  `build --behind` builds them; `push --behind` sends them on\n")
+		fmt.Println()
+	}
+
+	if len(asked.waiting) > 0 {
+		// The other half of "is anything out of date": a module behind its source says the
+		// catalogue is old, and this says a machine is — and only this one has somebody's change
+		// waiting inside it.
+		var told, never []string
+		for _, m := range asked.waiting {
+			if m.Never {
+				never = append(never, m.Node)
+				continue
+			}
+			told = append(told, m.Node)
+		}
+		if len(told) > 0 {
+			fmt.Printf("%d machine(s) are not running what the mesh would send them:\n  %s\n",
+				len(told), strings.Join(told, ", "))
+		}
+		if len(never) > 0 {
+			// Never told is not out of date. The remedy is the same push and the situation is
+			// not the same at all: nobody has ever asked this machine to be anything.
+			fmt.Printf("%d machine(s) have never been sent anything:\n  %s\n",
+				len(never), strings.Join(never, ", "))
+		}
+		fmt.Printf("\n  `push --behind` sends them\n\n")
+	}
+
+	if split := hostSplit(nodes); len(split) > 1 {
+		// **Before a declaration gains a field, every machine has to understand it** (novox/hq
+		// 04-ISSUES/087). A host refuses a declaration carrying a field it does not know, and refuses
+		// it whole, so every new field is a flag day: hosts first, then the controller. The mesh had
+		// no record of which host any machine ran, so that order was kept by somebody remembering it.
+		//
+		// **Disagreement, and deliberately not "behind".** A host reports its version as a commit, and
+		// commits have no order — the first version of this said "N machines run an older host" and
+		// named the three that were newer, because it compared two hashes as strings. What the mesh
+		// can say truthfully is that the machines do not all run the same host, and which machines
+		// hold which. Ordering needs a version that is ordered, and that is the host's to report.
+		versions := make([]string, 0, len(split))
+		for v := range split {
+			versions = append(versions, v)
+		}
+		sort.Strings(versions)
+		fmt.Printf("%d machine(s) do not all run the same host:\n", len(nodes))
+		for _, v := range versions {
+			sort.Strings(split[v])
+			fmt.Printf("  %-12s %s\n", v, strings.Join(split[v], ", "))
+		}
+		fmt.Printf("\n  a host refuses a declaration carrying a field it does not know, whole — so the\n" +
+			"  mesh may send only what every one of these understands. Which of them is newer is\n" +
+			"  not readable from a commit; that needs a version the host reports as ordered\n\n")
+	}
+
+	if len(asked.filtered) > 0 {
+		// A converged machine is filtered by the mesh alone, and the mesh says truthfully which
+		// (novox/hq ADR 0168). One that is not — a predecessor's chain still refusing, a found
+		// firewall in force again — is said here, and is not well.
+		machines := make([]string, 0, len(asked.filtered))
+		for name := range asked.filtered {
+			machines = append(machines, name)
+		}
+		sort.Strings(machines)
+		fmt.Printf("%d converged machine(s) are not filtered by the mesh alone:\n", len(machines))
+		for _, name := range machines {
+			f := asked.filtered[name]
+			if fw := f.FoundFirewall; fw != nil && fw.Active {
+				fmt.Printf("  %-12s the found firewall (%s) is in force again; the next apply retires it\n", name, fw.Kind)
+			}
+			for _, x := range f.Others() {
+				fmt.Printf("  %-12s %s (%s): %s\n", name, x.Where, x.Owner, x.Refuses)
+			}
+		}
+		fmt.Printf("\n  the mesh wrote none of these and removes none; `node show <node>` lists every filter with its owner\n\n")
+	}
+
+	if len(asked.untaken) > 0 {
+		// **Before the adopted line, and it breaks "all well".** An adopted machine is a state
+		// somebody chose and can leave alone; a module assigned to one and never taken is work
+		// outstanding that reads exactly like work finished. That reading is what stopped a
+		// predecessor's proxy on the strength of four green surfaces (novox/hq 04-ISSUES/125).
+		machines := make([]string, 0, len(asked.untaken))
+		for name := range asked.untaken {
+			machines = append(machines, name)
+		}
+		sort.Strings(machines)
+		total := 0
+		for _, held := range asked.untaken {
+			for _, n := range held {
+				total += n
+			}
+		}
+		fmt.Printf("%d resource(s) are held as found, because their module was assigned and never "+
+			"taken — so it is running none of what it declares:\n", total)
+		for _, name := range machines {
+			modules := make([]string, 0, len(asked.untaken[name]))
+			for m := range asked.untaken[name] {
+				modules = append(modules, m)
+			}
+			sort.Strings(modules)
+			parts := make([]string, 0, len(modules))
+			for _, m := range modules {
+				parts = append(parts, fmt.Sprintf("%s (%d)", m, asked.untaken[name][m]))
+			}
+			fmt.Printf("  %-12s %s\n", name, strings.Join(parts, ", "))
+		}
+		fmt.Printf("\n  `take <node> <module>` compares what runs against what it declares, and runs it\n\n")
+	}
+
+	if len(asked.unheld) > 0 {
+		// **Reported, and not refused yet** (novox/hq ADR 0207 §4). Each machine still resolves and
+		// is sent what it would be; this says which of its modules depend on a seat nothing there
+		// holds, until every machine has its holders and the switch makes it a refusal.
+		fmt.Printf("%d module dependenc(ies) on a seat nothing on the machine holds (unheld, ADR 0207):\n",
+			len(asked.unheld))
+		for _, u := range asked.unheld {
+			holders := "no module in the catalogue claims it yet"
+			if len(u.Holders) > 0 {
+				holders = "could be held by " + strings.Join(u.Holders, ", ")
+			}
+			fmt.Printf("  %-12s %-24s %-24s %s\n", u.Node, u.Module, u.Seat, holders)
+		}
+		fmt.Printf("\n  `assign <node> <holder>` meets it; reported until every machine has its holders, then refused\n\n")
+	}
+
+	if adopted := adoptedNodes(nodes); len(adopted) > 0 {
+		// Said, because nothing forces the flip: a node left adopted is visible here rather than
+		// read as converged (novox/hq ADR 0100). Not a fault, so it does not break "all well".
+		fmt.Printf("%d machine(s) adopted: %s\n", len(adopted), strings.Join(adopted, ", "))
+		fmt.Printf("\n  `converge <node>` previews the flip\n\n")
+	}
+
+	if asked.well() {
+		// Said plainly. "Nothing to report" and "nothing was checked" must never look the same,
+		// and getting here means every question was asked and answered.
+		fmt.Printf("%d machine(s), all doing what they were told, all heard from, running what "+
+			"the mesh would send them, and every module current with its source\n", len(nodes))
+		// **And what that sentence does not cover**, because for eleven hours it was true of a mesh
+		// in which no module could reach another (novox/hq 04-ISSUES/145). Every question above is
+		// about the relationship between the mesh and a machine — applied what it was sent, matches
+		// what would be sent, built from what the source has. None of them asks whether a module can
+		// reach what it requires, and the mesh composes every one of those grants itself.
+		//
+		// Said here rather than left to be inferred. A reader who acts on the line above is acting on
+		// "the machines are as the mesh described them", and the distance between that and "it works"
+		// is where the eleven hours went.
+		fmt.Printf("\n  That is the mesh and the machines agreeing. Nothing here dials a provision:\n" +
+			"  no grant the mesh composed has been tested, so a module unable to reach what it\n" +
+			"  requires would not appear above (04-ISSUES/145)\n")
+	}
+	return nil
+}
+
+// firstLine is as much of a failure as belongs in a list.
+func firstLine(s string) string {
+	if cut := strings.IndexByte(s, '\n'); cut >= 0 {
+		return strings.TrimSpace(s[:cut])
+	}
+	return strings.TrimSpace(s)
+}
+
+// builds keeps what a builder said, for the serving control plane.
+//
+// A type of its own rather than a method on the enrolment, because they are unrelated things
+// arriving on one queue and an implementation of one should not have to say anything about the
+// other.
+type builds struct {
+	inv  *inventory.Inventory
+	open *stores
+}
+
+// theThreeQuestions reads what anything answering "is the mesh alright" needs.
+//
+// **One reading, for every way of saying it** (novox/hq 03-DESIGN/01-to-be/11-a-board.md). There
+// are three now — a person's status, its JSON, and a page — and three implementations of "which
+// machine is not doing what it was told" would be three chances to disagree about it.
+//
+// The order is the design and not a convenience: is anything broken, is anything not answering, is
+// anything out of date. The first has consequences now, the second may, the third is a plan for
+// later — and anything that led with the third would bury the first.
+func theThreeQuestions(ctx context.Context, open *stores) (answers, error) {
+	inv := open.inventory
+	var out answers
+	var err error
+
+	out.wrong, err = inv.NotDoingWhatTheyWereTold(ctx)
+	if err != nil {
+		return answers{}, err
+	}
+	out.nodes, err = inv.Nodes(ctx)
+	if err != nil {
+		return answers{}, err
+	}
+	for _, n := range out.nodes {
+		// Never heard from, or not lately. Different from failing: a machine that says nothing
+		// may be new, switched off, or unreachable, and none of those is a machine that tried
+		// and could not.
+		if n.LastSeen.IsZero() || time.Since(n.LastSeen) > time.Hour {
+			out.quiet = append(out.quiet, n)
+		}
+	}
+	out.behind, err = inv.Behind(ctx)
+	if err != nil {
+		return answers{}, err
+	}
+	// And why any machine cannot be worked out at all, which is neither of the first two questions
+	// and is asked before them both in practice: a machine nothing can be computed for is not
+	// broken, not quiet and not behind, and every other answer here would call it well.
+	//
+	// Read through whoResolves, which is what the private network is built from, so this and the
+	// network agree about who could not be resolved rather than deciding it twice.
+	_, out.refused, err = whoResolves(ctx, open, overlay.Addressing)
+	if err != nil {
+		return answers{}, err
+	}
+	// And what each machine is holding rather than running, by the module that would run it. Read
+	// from what the machine itself last reported, not from what take-time computed: the machine is
+	// the only thing that knows what it found (novox/hq 04-ISSUES/125).
+	out.untaken, err = untakenModules(ctx, inv, out.nodes)
+	if err != nil {
+		return answers{}, err
+	}
+	// And which converged machines something other than the mesh filters (novox/hq ADR 0168), as
+	// each last reported — the account that was missing when a predecessor's chain refused what the
+	// mesh declared open for eleven hours (04-ISSUES/144, 145).
+	out.filtered, err = filteredMachines(ctx, inv, out.nodes)
+	if err != nil {
+		return answers{}, err
+	}
+	// And which machines run a module whose resources a seat nothing there holds applies (novox/hq
+	// ADR 0207). Each machine resolved again rather than threaded through whoResolves, whose answer
+	// the private network is built from and should say nothing else; a machine that does not
+	// resolve is already in refused, and is passed over here.
+	for _, n := range out.nodes {
+		plan, _, err := planFor(ctx, open, n.Name)
+		if err != nil {
+			if unresolvable(err) {
+				continue
+			}
+			return answers{}, err
+		}
+		out.unheld = append(out.unheld, plan.Unheld...)
+	}
+	// And every consumer a provider says it keeps failing (novox/hq ADR 0224). Read from what the
+	// providers announced: nothing else in the mesh knows whether a provision is being made.
+	out.failing, err = failingProviders(ctx, inv)
+	if err != nil {
+		return answers{}, err
+	}
+	out.plans, err = inv.RecentPlans(ctx, 5)
+	if err != nil {
+		return answers{}, err
+	}
+	// Whether the build seat takes work, for a plan waiting on it (novox/hq ADR 0219).
+	out.paused = buildSeatPause(ctx, inv, out.plans)
+
+	// And which machines are not running what the mesh would send them. The same question as a
+	// module being behind its source, one level down: that one says the catalogue is out of date,
+	// this one says a machine is — and only the second has anybody's change waiting in it.
+	//
+	// **One machine that cannot be resolved must not take the answer away from every other**
+	// (novox/hq 04-ISSUES/017's sibling). This reaches the private network, and a mesh whose hub
+	// is the blocked machine has no hub — which used to come back here as a refusal, so `status`
+	// said nothing at all and `status --json` emitted prose to stderr and no JSON anywhere. The
+	// reason is kept and reported as data; every question that does not depend on it is still
+	// answered.
+	would, err := wouldSend(ctx, open, out.nodes)
+	if err != nil {
+		out.network = err.Error()
+		would = map[string]string{}
+	}
+	out.waiting, err = inv.Waiting(ctx, would)
+	if err != nil {
+		return answers{}, err
+	}
+	out.reported, err = inv.LastReports(ctx)
+	if err != nil {
+		return answers{}, err
+	}
+	out.sources = map[string]inventory.Source{}
+	for module := range out.behind {
+		from, err := inv.SourceOf(ctx, module)
+		if err != nil {
+			return answers{}, err
+		}
+		out.sources[module] = from
+	}
+	return out, nil
+}
+
+// untakenModules is, per machine, each module whose resources that machine is holding as found, and
+// how many.
+//
+// **The machine's own account, not the mesh's.** An adopted node decides at apply time what it found
+// and reports it; the mesh's take-time listing is a different thing and was the one this command used
+// to have, which is why a module assigned after the listing showed nothing at all
+// (novox/hq 04-ISSUES/125).
+//
+// A machine that reports no holds contributes nothing, so a converged mesh answers an empty map and
+// the caller prints nothing.
+// filteredMachines is every converged machine not filtered by the mesh alone, with what it last said
+// filters it (novox/hq ADR 0168). An adopted machine keeps its found firewall by design and is not
+// counted; a machine that has not said is not said to be filtered by anything.
+func filteredMachines(ctx context.Context, inv *inventory.Inventory, nodes []inventory.Node) (
+	map[string]inventory.Filtering, error) {
+	out := map[string]inventory.Filtering{}
+	for _, n := range nodes {
+		if n.Adopted {
+			continue
+		}
+		f, err := inv.FilteringOf(ctx, n.Name)
+		if err != nil {
+			return nil, fmt.Errorf("what filters %s cannot be read: %w", n.Name, err)
+		}
+		if len(f.Filters) == 0 && f.FoundFirewall == nil {
+			continue
+		}
+		if !f.Alone() {
+			out[n.Name] = f
+		}
+	}
+	return out, nil
+}
+
+func untakenModules(ctx context.Context, inv *inventory.Inventory, nodes []inventory.Node) (
+	map[string]map[string]int, error) {
+
+	out := map[string]map[string]int{}
+	for _, n := range nodes {
+		said, err := inv.AdoptionOf(ctx, n.Name)
+		if err != nil {
+			// A machine whose record cannot be read is not a machine holding nothing. Said, because
+			// answering "nothing held" from a failed read is the shape this whole issue is about.
+			return nil, fmt.Errorf("what %s is holding cannot be read: %w", n.Name, err)
+		}
+		for _, h := range said.Held {
+			if h.Module == "" {
+				continue // a hold the mesh cannot attribute to a module has nothing to take
+			}
+			if out[n.Name] == nil {
+				out[n.Name] = map[string]int{}
+			}
+			out[n.Name][h.Module]++
+		}
+	}
+	return out, nil
+}
+
+// well is whether every question this command asks came back with nothing to say.
+//
+// Named, and in one place, because it is the sentence an operator acts on and it has been wrong
+// twice. It is deliberately NOT "nothing is broken": a machine holding what it found is not broken
+// and is not doing what it was told either.
+//
+// **A hold suppresses it; being adopted does not.** Adopted is a mode somebody chose and can leave
+// alone. A module assigned to a machine and never taken is a half-finished action with nothing left
+// to finish it — it runs none of what it declares, and "all doing what they were told" was true and
+// read as success for the whole of the edge cut-over outage (novox/hq 04-ISSUES/125).
+func (a answers) well() bool {
+	return len(a.wrong) == 0 && len(a.quiet) == 0 && len(a.behind) == 0 &&
+		len(a.waiting) == 0 && len(a.refused) == 0 && a.network == "" && len(a.untaken) == 0 &&
+		len(a.filtered) == 0 && len(a.unheld) == 0 && len(a.failing) == 0
+}
+
+// hostSplit is which machines report which host version, for every version more than one machine
+// could disagree about.
+//
+// **It does not say which is newer, because it cannot.** A host reports its version as a commit, and
+// commits have no order. The first version of this returned "the machines behind the newest" by
+// comparing versions as strings, and on the live mesh it named the three machines running the NEWER
+// host as the ones behind — an arbitrary lexicographic result presented as a fact
+// (novox/hq 04-ISSUES/087). A report that confidently says the opposite of the truth is worse than one
+// that says less, which is the whole subject of 04-ISSUES/145.
+//
+// So this answers what is checkable: who runs what. The reader sees the split and the mesh claims no
+// ordering. Ordering wants an ordered version, and that is the host's to report rather than this
+// function's to infer.
+//
+// Machines that have not reported a version are left out entirely: they are not a version, and
+// counting them as one would invent a disagreement. `node show` says per machine that it has not said.
+func hostSplit(nodes []inventory.Node) map[string][]string {
+	out := map[string][]string{}
+	for _, n := range nodes {
+		if n.HostVersion == "" {
+			continue
+		}
+		out[n.HostVersion] = append(out[n.HostVersion], n.Name)
+	}
+	if len(out) < 2 {
+		return nil // one version, or none reported: nothing to disagree about
+	}
+	return out
+}

cmd/mesh-controller/status_test.go

@@ -0,0 +1,121 @@
+package main
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+// **One machine's failure must never take the answer away from a reader asking about the others.**
+//
+// A blocked machine is not on the private network, and a mesh whose hub is that machine has no hub
+// — which came back through the reading as a refusal, so `status` printed nothing at all and
+// `status --json` emitted multi-line prose on stderr and not one byte of JSON. A machine-readable
+// interface that stops being machine-readable exactly when something is wrong is one nobody can
+// build an alarm on.
+func TestOneBlockedMachineDoesNotDestroyTheWholeDocument(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+
+	// Break the hub, using nothing but the command a person has.
+	one, two := rivals()
+	register(t, open, one)
+	register(t, open, two)
+	for _, m := range []string{"rival-one", "rival-two"} {
+		if _, err := assign(ctx, open, "anchor", m); err != nil && m == "rival-one" {
+			t.Fatal(err)
+		}
+	}
+
+	asked, err := theThreeQuestions(ctx, open)
+	if err != nil {
+		t.Fatalf("one blocked machine made the whole mesh unreadable: %v", err)
+	}
+	body, err := statusAsJSON(asked)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var parsed map[string]any
+	if err := json.Unmarshal(body, &parsed); err != nil {
+		t.Fatalf("what a script would read is not JSON: %v\n%s", err, body)
+	}
+
+	// The failure is in the document, as data, on the machine it belongs to.
+	unresolved, _ := parsed["unresolved"].([]any)
+	if len(unresolved) == 0 {
+		t.Fatalf("a machine that cannot be worked out is absent from the document:\n%s", body)
+	}
+	var found bool
+	for _, row := range unresolved {
+		entry, _ := row.(map[string]any)
+		if entry["node"] != "anchor" {
+			continue
+		}
+		found = true
+		problem, _ := entry["problem"].(string)
+		if !strings.Contains(problem, "the-seat") {
+			t.Errorf("the machine's problem is not its own words: %q", problem)
+		}
+	}
+	if !found {
+		t.Fatalf("the blocked machine is not the one named:\n%s", body)
+	}
+	// And the mesh is still counted, so a reader can tell "none blocked" from "none at all".
+	if parsed["machines"] != float64(2) {
+		t.Errorf("the rest of the document did not survive: %v", parsed["machines"])
+	}
+}
+
+// A well mesh carries the field as an empty list, not as nothing: a reader distinguishing "none
+// blocked" from "this field is missing" would have to handle both, and null is the one that gets
+// forgotten.
+func TestAWellMeshCarriesAnEmptyUnresolvedList(t *testing.T) {
+	open := aMesh(t)
+	asked, err := theThreeQuestions(t.Context(), open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	body, err := statusAsJSON(asked)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var parsed map[string]any
+	if err := json.Unmarshal(body, &parsed); err != nil {
+		t.Fatal(err)
+	}
+	list, ok := parsed["unresolved"].([]any)
+	if !ok {
+		t.Fatalf("\"unresolved\" is %T, not a list:\n%s", parsed["unresolved"], body)
+	}
+	if len(list) != 0 {
+		t.Fatalf("a well mesh reports blocked machines: %v", list)
+	}
+	if _, said := parsed["network"]; said {
+		t.Errorf("a well mesh says the network could not be computed: %v", parsed["network"])
+	}
+}
+
+// And the page says it too, from the same reading. A board that renders "all well" over a mesh
+// where nothing can be sent anywhere is worse than a board that is down.
+func TestThePageSaysWhichMachinesCannotBeWorkedOut(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	one, two := rivals()
+	register(t, open, one)
+	register(t, open, two)
+	for _, m := range []string{"rival-one", "rival-two"} {
+		if _, err := assign(ctx, open, "anchor", m); err != nil && m == "rival-one" {
+			t.Fatal(err)
+		}
+	}
+	asked, err := theThreeQuestions(ctx, open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	rendered := render(t, viewOf(asked))
+	for _, want := range []string{"Can everything be worked out?", "anchor", "the-seat"} {
+		if !strings.Contains(rendered, want) {
+			t.Fatalf("the page does not say %q:\n%s", want, rendered)
+		}
+	}
+}

cmd/mesh-controller/stores.go

@@ -0,0 +1,215 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/identity"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/licences"
+	"github.com/novox/mesh-controller/internal/store"
+)
+
+// reaching each context's store, which no other context may touch.
+//
+// Split out of main.go, which had reached 2,769 lines because appending was always the
+// cheapest next step. That is how novox/hq ADR 0001 records `hal/sdk` reaching 34,636:
+// nothing in it was wrong, and no one edit was the one that should have been a new file.
+
+// migrate brings every held context's schema up to date.
+//
+// Reported per context and per migration, because this runs during a bootstrap on a machine with
+// nothing else on it — the output is the only account of what happened, and "migrated" is not one.
+func migrate(ctx context.Context) error {
+	for _, c := range held {
+		migrations, err := c.migrations()
+		if err != nil {
+			return err
+		}
+
+		s, err := store.Open(ctx, c.name)
+		if err != nil {
+			return err
+		}
+		defer s.Close()
+
+		// The bootstrap raises PostgreSQL moments before this runs, and a container that is
+		// running is not a database that will answer — a distinction this project has already
+		// paid for once, when a crash-looping database reported itself as up between restarts.
+		if err := s.Ready(ctx, 60*time.Second); err != nil {
+			return err
+		}
+
+		done, err := s.Migrate(ctx, migrations)
+		for _, m := range done {
+			fmt.Printf("%s: applied %04d-%s\n", c.name, m.Number, m.Name)
+		}
+		if err != nil {
+			return err
+		}
+		if len(done) == 0 {
+			applied, err := s.AppliedMigrations(ctx)
+			if err != nil {
+				return err
+			}
+			fmt.Printf("%s: already up to date — %d migration(s)\n", c.name, len(applied))
+		}
+	}
+
+	// The modules the control plane ships with itself. Recorded here rather than by hand, because
+	// a mesh whose own private network is missing from the catalogue would have nothing to assign
+	// and no way to say why.
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+	for _, m := range provided {
+		if err := inv.Provide(ctx, m); err != nil {
+			return err
+		}
+		fmt.Printf("provided %s\n", m.Module)
+	}
+	// And what an earlier release shipped and this one does not goes (novox/hq ADR 0226) — unless a
+	// machine still has it, which is said rather than overridden.
+	var shipped []string
+	for _, m := range provided {
+		shipped = append(shipped, m.Module)
+	}
+	retired, kept, err := inv.RetireUnshipped(ctx, shipped)
+	if err != nil {
+		return err
+	}
+	for _, name := range retired {
+		fmt.Printf("retired %s: the control plane no longer ships it\n", name)
+	}
+	for name, why := range kept {
+		fmt.Printf("kept %s, which the control plane no longer ships: %s\n", name, why)
+	}
+	// The seats the mesh ships with, into the table that now holds the set (novox/hq ADR 0122).
+	// Idempotent: fills an empty table on first boot, adds a seat a release ships, and leaves an
+	// operator's changes in the table as they are.
+	added, err := inv.SeedSeats(ctx, catalogue.DefaultSeats())
+	if err != nil {
+		return err
+	}
+	fmt.Printf("seeded %d seat(s)\n", added)
+	return nil
+}
+
+// openInventory connects and waits, the way every command that touches it needs to.
+func openInventory(ctx context.Context) (*inventory.Inventory, error) {
+	inv, err := inventory.Open(ctx)
+	if err != nil {
+		return nil, err
+	}
+	if err := inv.Ready(ctx, 30*time.Second); err != nil {
+		inv.Close()
+		return nil, err
+	}
+	// Load the seat set from the store, so the control plane reads the set as data rather than as
+	// the slice it was compiled with (novox/hq ADR 0122). A store not yet seeded — or one whose
+	// seat table a migration has not reached — returns nothing, and UseSeats leaves the compiled
+	// defaults in force: the set is never emptied by a read that found nothing, which would refuse
+	// every claim. So this can only ever replace the defaults with what the mesh actually holds.
+	if seats, err := inv.Seats(ctx); err == nil {
+		catalogue.UseSeats(seats)
+	}
+	// And the former names, so a reference to a seat's old name resolves after a rename (ADR 0122).
+	if aliases, err := inv.Aliases(ctx); err == nil {
+		catalogue.UseAliases(aliases)
+	}
+	return inv, nil
+}
+
+func openIdentity(ctx context.Context) (*identity.Identity, error) {
+	ident, err := identity.Open(ctx)
+	if err != nil {
+		return nil, err
+	}
+	if err := ident.Ready(ctx, 30*time.Second); err != nil {
+		ident.Close()
+		return nil, err
+	}
+	return ident, nil
+}
+
+// openLicences connects to the context that holds which model access exists and who may use it.
+func openLicences(ctx context.Context) (*licences.Licences, error) {
+	held, err := licences.Open(ctx)
+	if err != nil {
+		return nil, err
+	}
+	if err := held.Ready(ctx, 30*time.Second); err != nil {
+		held.Close()
+		return nil, err
+	}
+	return held, nil
+}
+
+// stores is what one command has open.
+//
+// **Opened once, not once per machine.** Working out what a machine should be reaches the identity
+// context for its certificate and the licence context for its model access, and both were opened —
+// and waited on — inside functions called for every node in a push. Two machines hid it; fifty
+// would be fifty connect-and-wait cycles for data that does not change while the push runs.
+//
+// Each is opened on first use rather than up front, because most commands need one context and
+// paying to reach three would be the same waste from the other side.
+type stores struct {
+	inventory *inventory.Inventory
+	identity  *identity.Identity
+	licences  *licences.Licences
+}
+
+// open connects to the inventory, which every command that touches the mesh needs.
+func openStores(ctx context.Context) (*stores, error) {
+	inv, err := openInventory(ctx)
+	if err != nil {
+		return nil, err
+	}
+	return &stores{inventory: inv}, nil
+}
+
+// Identity is this control plane's own identity context, opened if it has not been.
+func (h *stores) Identity(ctx context.Context) (*identity.Identity, error) {
+	if h.identity != nil {
+		return h.identity, nil
+	}
+	opened, err := openIdentity(ctx)
+	if err != nil {
+		return nil, err
+	}
+	h.identity = opened
+	return opened, nil
+}
+
+// Licences is the context holding model access, opened if it has not been.
+func (h *stores) Licences(ctx context.Context) (*licences.Licences, error) {
+	if h.licences != nil {
+		return h.licences, nil
+	}
+	opened, err := openLicences(ctx)
+	if err != nil {
+		return nil, err
+	}
+	h.licences = opened
+	return opened, nil
+}
+
+// Close lets go of everything that was opened, in any order: they are separate connections to
+// separate databases and none of them knows about the others.
+func (h *stores) Close() {
+	if h.licences != nil {
+		h.licences.Close()
+	}
+	if h.identity != nil {
+		h.identity.Close()
+	}
+	if h.inventory != nil {
+		h.inventory.Close()
+	}
+}

cmd/mesh-controller/supersede_test.go

@@ -0,0 +1,95 @@
+package main
+
+import (
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// novox/hq issue 254, ADR 0218: a newer plan takes over what the older open plans of its repository
+// and branch had not built, and closes them as superseded; another repository's plan, another
+// branch's, and a plan made after it are left alone.
+func TestANewerPlanSupersedesTheOlderOpenPlansOfItsRepository(t *testing.T) {
+	at := time.Date(2026, 10, 5, 12, 0, 0, 0, time.UTC)
+	sent := at.Add(time.Minute)
+	plan := func(id, repository, branch string, created time.Time, modules map[string]*inventory.PlanModule) inventory.Plan {
+		return inventory.Plan{ID: id, Repository: repository, Branch: branch, Commit: id + "-commit",
+			Created: created, State: inventory.PlanRolling, Modules: modules}
+	}
+	older := plan("plan-1", "novox/mesh-catalog", "main", at, map[string]*inventory.PlanModule{
+		"gitea":    {State: "built", SentAt: &sent}, // done with: stays done
+		"keycloak": {State: "asked"},                // asked, not answered: folded
+		"plex":     {},                              // not yet asked: folded
+		"agent":    {State: "built"},                // built, rolls out, not sent: folded
+		"notes":    {State: "built"},                // built, records: nothing to send
+	})
+	stuck := plan("plan-0", "Novox/Mesh-Catalog", "", at.Add(-time.Hour), map[string]*inventory.PlanModule{
+		"runtime": {State: "asked"},
+	})
+	other := plan("plan-2", "novox/mesh-controller", "main", at, map[string]*inventory.PlanModule{"mesh-controller": {}})
+	release := plan("plan-3", "novox/mesh-catalog", "release", at, map[string]*inventory.PlanModule{"lemurs": {}})
+	later := plan("plan-5", "novox/mesh-catalog", "main", at.Add(2*time.Hour), map[string]*inventory.PlanModule{"later": {}})
+	done := plan("plan-6", "novox/mesh-catalog", "main", at, map[string]*inventory.PlanModule{"finished": {}})
+	done.State = inventory.PlanDone
+
+	newer := plan("plan-4", "novox/mesh-catalog", "main", at.Add(time.Hour), nil)
+	newer.Commit = "97b1b2b0c0ffee"
+	rollsOut := func(m string) bool { return m != "notes" }
+	folded, closed := supersededBy(newer, []inventory.Plan{stuck, older, other, release, later, done, newer}, rollsOut)
+
+	if want := []string{"agent", "keycloak", "plex", "runtime"}; !reflect.DeepEqual(folded, want) {
+		t.Fatalf("folded %v, wanted %v", folded, want)
+	}
+	var ids []string
+	for _, p := range closed {
+		ids = append(ids, p.ID)
+		if p.State != inventory.PlanSuperseded || p.Open() {
+			t.Errorf("%s was left %s", p.ID, p.State)
+		}
+		if !strings.Contains(p.Note, "plan-4") || !strings.Contains(p.Note, "97b1b2b0") {
+			t.Errorf("%s does not name the plan that superseded it: %q", p.ID, p.Note)
+		}
+	}
+	if want := []string{"plan-0", "plan-1"}; !reflect.DeepEqual(ids, want) {
+		t.Fatalf("superseded %v, wanted %v — another repository, another branch, a later plan and a "+
+			"finished one are left alone", ids, want)
+	}
+	if other.State != inventory.PlanRolling {
+		t.Fatal("the plan handed in was changed in place")
+	}
+	if line := planLine(closed[1], time.Now()); !strings.Contains(line, "superseded") {
+		t.Fatalf("a superseded plan reads %q", line)
+	}
+}
+
+// novox/hq issue 254: a person closes a plan that will not move again, by its id.
+func TestAPersonClosesAStuckPlan(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	stuck := inventory.Plan{ID: "plan-97b1b2b", Repository: "novox/mesh-catalog", Commit: "97b1b2b",
+		Created: time.Now().UTC(), State: inventory.PlanRolling, Tier: 1, Tiers: [][]string{{"a"}, {"b"}},
+		Modules: map[string]*inventory.PlanModule{"a": {State: "built"}, "b": {}}}
+	if err := open.inventory.SavePlan(ctx, stuck); err != nil {
+		t.Fatal(err)
+	}
+	if err := plansCommand(ctx, []string{"close", stuck.ID}); err != nil {
+		t.Fatal(err)
+	}
+	closed, err := open.inventory.PlanByID(ctx, stuck.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if closed.State != inventory.PlanFailed || !strings.Contains(closed.Note, "closed by hand") {
+		t.Fatalf("the plan was left %s: %q", closed.State, closed.Note)
+	}
+	if err := plansCommand(ctx, []string{"close", stuck.ID}); err == nil {
+		t.Fatal("a plan already closed was closed again")
+	}
+	if argv, err := argvFor("plans", map[string]any{"close": stuck.ID}); err != nil ||
+		!reflect.DeepEqual(argv, []string{"plans", "close", stuck.ID}) {
+		t.Fatalf("the seat's verb does not close a plan: %v %v", argv, err)
+	}
+}

cmd/mesh-controller/take_preview_test.go

@@ -0,0 +1,143 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// What the forge's take compares, as a machine would report it.
+func aForgeComparison() comparison {
+	return comparison{reported: inventory.Adoption{
+		Firewall: "ufw",
+		Held: []inventory.Held{
+			{ID: "forge.server", Module: "forge", Kind: "container", Target: "forge", Facts: map[string]any{
+				"image": "forge:1.27.3", "image_created": "2026-09-17T10:00:00Z",
+				"declared_image": "forge:1.22.6", "declared_image_created": "2026-08-20T10:00:00Z", "downgrade": true,
+				"networks": map[string]any{"predecessor_default": []any{"office", "db"}},
+				"ports":    []any{"3000/tcp>0.0.0.0:3000"}, "declared_ports": []any{"3000:3000"},
+			}},
+			{ID: "forge.config", Module: "forge", Kind: "file", Target: "/etc/forge/app.ini", Kept: "/var/lib/mesh/kept/app.ini",
+				Facts: map[string]any{"differs": true, "difference": []any{"- private scope: local", "+ upstream: public"}}},
+			{ID: "other.server", Module: "other", Kind: "container", Target: "other"},
+		},
+		Reachable: []inventory.Reach{
+			{Protocol: "tcp", Address: "0.0.0.0", Port: 3000, By: "forge", Published: true, ContainerPort: 3000},
+			{Protocol: "tcp", Address: "0.0.0.0", Port: 22, By: "sshd"},
+		},
+	}}
+}
+
+// A take is a comparison (novox/hq ADR 0163): the preview puts what runs beside what the module
+// declares, and an older image or a differing file refuses unless named.
+func TestATakePreviewsTheComparisonAndRefusesWhatIsNotNamed(t *testing.T) {
+	c := aForgeComparison()
+	preview, refusals, saw := comparisonOf("forge", c, takeOptions{})
+	for _, want := range []string{"runs forge:1.27.3 (made 2026-09-17)", "declares forge:1.22.6 (made 2026-08-20)", "DOWNGRADE",
+		"on the network predecessor_default with office, db", "will not once it moves to the module's own network",
+		"publishes 3000/tcp>0.0.0.0:3000; the module declares 3000:3000",
+		// How far the port reaches now, as the machine reported it (rule 1).
+		"reachable now at 0.0.0.0:3000 (tcp, container port 3000), behind the found firewall (ufw)",
+		"- private scope: local", "original kept at /var/lib/mesh/kept/app.ini"} {
+		if !strings.Contains(preview, want) {
+			t.Errorf("the preview lacks %q:\n%s", want, preview)
+		}
+	}
+	if strings.Contains(preview, "other") {
+		t.Errorf("another module's held things are in the preview:\n%s", preview)
+	}
+	if len(refusals) != 2 || !strings.Contains(refusals[0], "--downgrade") || !strings.Contains(refusals[1], "--replace /etc/forge/app.ini") {
+		t.Fatalf("the downgrade and the differing file refuse, each naming its override: %v", refusals)
+	}
+	if len(saw) != 12 {
+		t.Fatalf("the preview's digest is %q", saw)
+	}
+	// Named, they pass.
+	if _, refusals, _ := comparisonOf("forge", c, takeOptions{Downgrade: true, Replace: map[string]bool{"/etc/forge/app.ini": true}}); len(refusals) != 0 {
+		t.Fatalf("named differences still refused: %v", refusals)
+	}
+	if _, refusals, _ := comparisonOf("forge", c, takeOptions{Downgrade: true, Replace: map[string]bool{"*": true}}); len(refusals) != 0 {
+		t.Fatalf("replace * did not cover the file: %v", refusals)
+	}
+	// A held thing with no facts yet — a host older than this — refuses nothing and says what it can.
+	if preview, refusals, _ := comparisonOf("other", c, takeOptions{}); len(refusals) != 0 || !strings.Contains(preview, "container other") {
+		t.Fatalf("a factless hold: %q %v", preview, refusals)
+	}
+	// The digest is of what the preview says: a fact changing changes it.
+	c.reported.Held[0].Facts["image"] = "forge:1.27.4"
+	if _, _, again := comparisonOf("forge", c, takeOptions{}); again == saw {
+		t.Fatal("the found image changed and the digest did not")
+	}
+}
+
+// A secret the mesh minted for a service whose data was found refuses: the running service already
+// has a value (rule 2). Accepted, it is carried in; `--mint` says the service shall take the new one.
+func TestAMintedSecretForFoundDataRefusesUnlessAcceptedOrMinted(t *testing.T) {
+	c := aForgeComparison()
+	c.secrets = []inventory.SecretState{
+		{Name: "admin", Origin: inventory.OriginMade},
+		{Name: "postgres-database", Origin: inventory.OriginMade, Provider: "anchor"},
+		{Name: "broker", Origin: inventory.OriginAccepted},
+	}
+	preview, refusals, _ := comparisonOf("forge", c, takeOptions{Downgrade: true, Replace: map[string]bool{"*": true}})
+	for _, want := range []string{
+		"own secret admin: MINTED by the mesh and not accepted",
+		"secret from anchor postgres-database: MINTED by the mesh and not accepted",
+		"own secret broker: accepted from a person, carried in as it is",
+	} {
+		if !strings.Contains(preview, want) {
+			t.Errorf("the preview lacks %q:\n%s", want, preview)
+		}
+	}
+	if len(refusals) != 2 {
+		t.Fatalf("two minted secrets refuse: %v", refusals)
+	}
+	if !strings.Contains(refusals[0], "`secret accept <node> forge admin`") || !strings.Contains(refusals[0], "`--mint admin`") {
+		t.Errorf("the own secret's refusal names accepting it and minting it: %s", refusals[0])
+	}
+	if !strings.Contains(refusals[1], "`secret accept <node> forge postgres-database --provider anchor`") {
+		t.Errorf("the required secret's refusal names its provider: %s", refusals[1])
+	}
+	preview, refusals, _ = comparisonOf("forge", c, takeOptions{Downgrade: true, Replace: map[string]bool{"*": true},
+		Mint: map[string]bool{"admin": true, "postgres-database": true}})
+	if len(refusals) != 0 || !strings.Contains(preview, "admin: minted by the mesh; the service takes the new value, as --mint said") {
+		t.Fatalf("--mint did not pass the minted secrets: %v\n%s", refusals, preview)
+	}
+	// With no found data — only a file held — the service has no value of its own, and a minted
+	// secret is simply said.
+	c.reported.Held = c.reported.Held[1:2]
+	if _, refusals, _ := comparisonOf("forge", c, takeOptions{Replace: map[string]bool{"*": true}}); len(refusals) != 0 {
+		t.Fatalf("a minted secret refused with no data found: %v", refusals)
+	}
+}
+
+// A found network a per-machine setting keeps is named in the preview (rule 4), and the module's
+// settings are said with where each came from, composed or not (rules 1 and 6).
+func TestTheKeptNetworkAndTheSettingsAreInThePreview(t *testing.T) {
+	c := aForgeComparison()
+	c.keeps = map[string][]string{"forge.server": {"predecessor_default"}}
+	c.layers = []catalogue.Layer{
+		{From: catalogue.MeshWideLayer, Values: map[string]any{"site": "x"}},
+		{From: "anchor", Values: map[string]any{catalogue.NetworksSetting: map[string]any{"server": []any{"predecessor_default"}}}},
+	}
+	preview, _, _ := comparisonOf("forge", c, takeOptions{Downgrade: true, Replace: map[string]bool{"*": true}})
+	for _, want := range []string{
+		"on the network predecessor_default with office, db — kept by this machine's setting, so they still reach it by name once taken",
+		"settings from the mesh: site",
+		"settings from anchor: networks",
+	} {
+		if !strings.Contains(preview, want) {
+			t.Errorf("the preview lacks %q:\n%s", want, preview)
+		}
+	}
+	if strings.Contains(preview, "will not once it moves") {
+		t.Errorf("a kept network is still said to be lost:\n%s", preview)
+	}
+	c.settingsRefused = "forge: ports is a { port: machine-port } map"
+	preview, _, _ = comparisonOf("forge", c, takeOptions{Downgrade: true, Replace: map[string]bool{"*": true}})
+	if !strings.Contains(preview, "SETTINGS DO NOT COMPOSE with the module's definition, so the push leaves it out: forge: ports") {
+		t.Errorf("settings that cannot compose are not said:\n%s", preview)
+	}
+}

cmd/mesh-controller/unheld_report_test.go

@@ -0,0 +1,108 @@
+package main
+
+import (
+	"bytes"
+	"os"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// An act says what it changed about the node it acted on, and nothing about the rest of the mesh
+// (novox/hq ADR 0207): after the seat dependencies shipped, every `push <node>` and `assign` printed
+// every node's unmet dependencies, a hundred lines around the one about the module just assigned.
+
+func aContainer(name string) catalogue.Manifest {
+	return catalogue.Manifest{Module: name, Version: "1",
+		Resources: []map[string]any{{"id": name, "type": "container", "image": name}}}
+}
+
+func TestAnAssignmentSaysOnlyWhatItChangedOnItsOwnNode(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	register(t, open, aContainer("web"))
+	register(t, open, aContainer("db"))
+	// anchor already lacks a runtime for db: true, and not this act's to say.
+	if _, err := open.inventory.Assign(ctx, "anchor", "db"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := open.inventory.Assign(ctx, "laptop", "db"); err != nil {
+		t.Fatal(err)
+	}
+	said, err := assign(ctx, open, "laptop", "web")
+	if err != nil {
+		t.Fatalf("%v\n%s", err, said)
+	}
+	if !strings.Contains(said, "web on laptop depends on "+catalogue.ContainerRuntimeSeat) {
+		t.Errorf("the assignment does not say what the module it assigned depends on:\n%s", said)
+	}
+	for _, not := range []string{"db on laptop", "db on anchor", "anchor:"} {
+		if strings.Contains(said, not) {
+			t.Errorf("the assignment says %q, which it did not change:\n%s", not, said)
+		}
+	}
+}
+
+func TestAnAssignmentThatMeetsADependencySaysSo(t *testing.T) {
+	shelf := map[string]catalogue.Manifest{
+		"web": aContainer("web"),
+		"docker": {Module: "docker", Claims: []catalogue.Claim{{Name: catalogue.ContainerRuntimeSeat}},
+			Resources: []map[string]any{{"id": "d", "type": "container", "image": "dind"}}},
+	}
+	lines := unheldChange(shelf, "laptop", []string{"web"}, []string{"web", "docker"})
+	if len(lines) != 1 || !strings.Contains(lines[0], "web on laptop now has "+catalogue.ContainerRuntimeSeat+" held") {
+		t.Errorf("meeting a dependency said %v", lines)
+	}
+	// Taking the dependent off says nothing: the dependency went with its module.
+	if lines := unheldChange(shelf, "laptop", []string{"web"}, nil); len(lines) != 0 {
+		t.Errorf("unassigning the dependent said %v", lines)
+	}
+}
+
+func TestAPushSaysANamedNodesDependenciesAndCountsTheRest(t *testing.T) {
+	unheld := map[string][]catalogue.Unheld{
+		"anchor": {{Node: "anchor", Module: "db", Seat: catalogue.ContainerRuntimeSeat}},
+		"laptop": {{Node: "laptop", Module: "web", Seat: catalogue.ContainerRuntimeSeat},
+			{Node: "laptop", Module: "sshd", Seat: catalogue.ServiceManagerSeat}},
+	}
+	var named bytes.Buffer
+	reportUnheldPushed(&named, true, []string{"laptop"}, unheld)
+	got := named.String()
+	if !strings.Contains(got, "laptop has 2 unmet") || !strings.Contains(got, "web on laptop") ||
+		!strings.Contains(got, "sshd on laptop") || strings.Contains(got, "anchor") {
+		t.Errorf("a named push said:\n%s", got)
+	}
+	var all bytes.Buffer
+	reportUnheldPushed(&all, false, []string{"anchor", "laptop", "quiet"}, unheld)
+	want := "anchor: 1 unmet seat dependenc(ies) — see `status`\nlaptop: 2 unmet seat dependenc(ies) — see `status`\n"
+	if all.String() != want {
+		t.Errorf("a push to every node said:\n%s\nwant\n%s", all.String(), want)
+	}
+}
+
+func TestOnlyTheServingControllerLogsEachChange(t *testing.T) {
+	read := func(f func()) string {
+		old := os.Stderr
+		r, w, _ := os.Pipe()
+		os.Stderr = w
+		f()
+		_ = w.Close()
+		os.Stderr = old
+		var b bytes.Buffer
+		_, _ = b.ReadFrom(r)
+		return b.String()
+	}
+	u := []catalogue.Unheld{{Node: "n1", Module: "web", Seat: catalogue.ContainerRuntimeSeat}}
+	if got := read(func() { logUnheld("n1", u) }); got != "" {
+		t.Errorf("a command logged:\n%s", got)
+	}
+	logUnheldChanges = true
+	defer func() { logUnheldChanges = false }()
+	if got := read(func() { logUnheld("n1", u) }); !strings.Contains(got, "web on n1") {
+		t.Errorf("the serving controller did not log a change:\n%s", got)
+	}
+	if got := read(func() { logUnheld("n1", u) }); got != "" {
+		t.Errorf("an unchanged report was logged again:\n%s", got)
+	}
+}

cmd/mesh-controller/untaken_test.go

@@ -0,0 +1,122 @@
+package main
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// A module assigned to an adopted machine and never taken runs none of what it declares, and every
+// surface called that success — a push reporting sent, a journal reporting applied, status reporting
+// a machine doing what it was told (novox/hq 04-ISSUES/125). The holds were only ever in the
+// machine's own state file.
+
+// heldOn makes a machine report that it is holding resources for a module, the way an adopted node
+// does after an apply.
+func heldOn(t *testing.T, open *stores, node, module string, ids ...string) {
+	t.Helper()
+	record, err := open.inventory.NodeByName(t.Context(), node)
+	if err != nil {
+		t.Fatal(err)
+	}
+	held := make([]inventory.Held, 0, len(ids))
+	for _, id := range ids {
+		held = append(held, inventory.Held{ID: id, Module: module, Kind: "container", Target: id})
+	}
+	if err := open.inventory.RecordAdoption(t.Context(), record.ID, held, "ufw", nil); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestStatusNamesAModuleHeldBecauseNothingTookIt(t *testing.T) {
+	open := aMesh(t)
+	heldOn(t, open, "anchor", "route-proxy", "ca", "certs", "server")
+
+	asked, err := theThreeQuestions(t.Context(), open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := asked.untaken["anchor"]["route-proxy"]; got != 3 {
+		t.Fatalf("status counted %d resources held for route-proxy, wanted 3", got)
+	}
+}
+
+func TestAHeldModuleStopsTheMeshReadingAsWell(t *testing.T) {
+	// The whole of the fault. "all doing what they were told" was true throughout the outage, and
+	// true is not the same as safe to act on: the machine was doing what it was told, and what it
+	// was told had not started. Asserted against the production condition, not a copy of it.
+	quiet := answers{}
+	if !quiet.well() {
+		t.Fatal("a mesh with nothing to say does not read as well, so nothing below means anything")
+	}
+	holding := answers{untaken: map[string]map[string]int{"anchor": {"route-proxy": 3}}}
+	if holding.well() {
+		t.Fatal("a machine holding a module's resources still reads as doing what it was told, " +
+			"which is the sentence that cost every public name on the machine")
+	}
+	// And being adopted does not suppress it: that is a mode somebody chose, not work outstanding.
+	// Kept as an assertion so the difference between the two is deliberate rather than incidental.
+	if !quiet.well() {
+		t.Fatal("the well condition is not stable")
+	}
+}
+
+func TestAHeldModuleIsFoundFromWhatTheMachineReported(t *testing.T) {
+	// End to end through the store, so the condition above is reached by real data and not only by
+	// a constructed value: the machine reports, the mesh records, status asks.
+	open := aMesh(t)
+	heldOn(t, open, "anchor", "route-proxy", "ca", "server")
+
+	asked, err := theThreeQuestions(t.Context(), open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(asked.untaken) == 0 {
+		t.Fatal("what the machine reported holding did not reach status")
+	}
+	if asked.well() {
+		t.Fatal("a mesh whose machine reported holds reads as well")
+	}
+}
+
+func TestTheJSONStatusCarriesWhatIsHeldAndForWhichModule(t *testing.T) {
+	open := aMesh(t)
+	heldOn(t, open, "anchor", "route-proxy", "ca", "certs")
+
+	asked, err := theThreeQuestions(t.Context(), open)
+	if err != nil {
+		t.Fatal(err)
+	}
+	body, err := statusAsJSON(asked)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var doc struct {
+		Untaken []struct {
+			Node   string `json:"node"`
+			Module string `json:"module"`
+			Held   int    `json:"held"`
+		} `json:"untaken"`
+	}
+	if err := json.Unmarshal(body, &doc); err != nil {
+		t.Fatal(err)
+	}
+	if len(doc.Untaken) != 1 {
+		t.Fatalf("the document carries %d untaken rows, wanted 1: %s", len(doc.Untaken), body)
+	}
+	row := doc.Untaken[0]
+	if row.Node != "anchor" || row.Module != "route-proxy" || row.Held != 2 {
+		t.Fatalf("the row is %+v, wanted anchor/route-proxy/2", row)
+	}
+	// Absent rather than empty when nothing is held, so a well mesh's document does not carry a
+	// field a reader has to interpret.
+	clean, err := statusAsJSON(answers{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(string(clean), "untaken") {
+		t.Fatalf("a mesh holding nothing still names untaken: %s", clean)
+	}
+}

cmd/mesh-controller/upgrades.go

@@ -0,0 +1,751 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"path"
+	"regexp"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// following acts on what the catalogue announces.
+//
+// **It holds the stores, not a copy of the decision.** What to do about an upgrade is read when
+// one arrives, so changing it takes effect on the next upgrade rather than on the next restart of
+// the control plane.
+type following struct{ open *stores }
+
+// Upgraded sends the machines running a module the version the catalogue now considers current —
+// or records that they are behind, which is the default and needs no record.
+//
+// **Recording is not a second code path.** A machine that is not running what the mesh would send
+// it is already something the mesh notices and reports; that is what `status` and `push --behind`
+// are built on. So "record it" is the absence of an action, and the only thing this has to decide
+// is whether to act.
+func (f following) Upgraded(ctx context.Context, u link.Upgraded) error {
+	inv := f.open.inventory
+
+	// The store read first, and an outage there said as one, so the announcement is held and asked
+	// again (novox/hq issue 083). Only here: a push that fails further down is not asked again.
+	decision, err := inv.UpgradeOf(ctx, u.Module)
+	if err != nil {
+		return notNow(err)
+	}
+	on, err := inv.Running(ctx, u.Module)
+	if err != nil {
+		return notNow(err)
+	}
+	if len(on) == 0 {
+		fmt.Printf("%s moved to %s; no machine runs it\n", u.Module, shortCommit(u.Commit))
+		return nil
+	}
+	if !decision.RollOut {
+		// Named rather than counted, and said even though nothing happens: an upgrade that was
+		// deliberately not rolled out and an upgrade that was never noticed look identical in a
+		// log that only speaks when it acts.
+		fmt.Printf("%s moved to %s; %s %s behind it, and this mesh records upgrades rather than "+
+			"rolling them out — `push --behind` when you want them\n",
+			u.Module, shortCommit(u.Commit), readableList(on), isAre(len(on)))
+		return nil
+	}
+
+	// **A plan that holds the module rolls it out, and this does not** (novox/hq issue 249, ADR
+	// 0218). A merge's plan builds the module and sends it one machine first, the rest once that one
+	// has applied it; this announcement arrives as the build registers, and sending here too — one
+	// machine after another without waiting for any to apply — put the new bundle on every machine in
+	// the same minute, whatever the plan was waiting for. A move no plan answers (a build asked by
+	// hand) is still this handler's.
+	if plans, err := inv.OpenPlans(ctx); err != nil {
+		return notNow(err)
+	} else if id := rolledOutByAPlan(plans, u.Module); id != "" {
+		// Said with its remedy: a plan that ends without sending it — failed, or closed by hand — leaves
+		// these machines behind, which `status` lists and `push --behind` sends (novox/hq issue 249).
+		fmt.Printf("%s moved to %s; %s rolls it out to %s — if that plan ends without sending it, "+
+			"`status` lists them as behind and `push --behind` sends it\n",
+			u.Module, shortCommit(u.Commit), id, readableList(on))
+		return nil
+	}
+	if decision.Together {
+		fmt.Printf("%s moved to %s; sending %s together\n",
+			u.Module, shortCommit(u.Commit), readableList(on))
+		return askAgainOnGrants(sendTo(ctx, f.open, on))
+	}
+	// One at a time, and stopping at the first that fails.
+	//
+	// **Stopping is the point.** The machines are done one after another precisely so that a
+	// version that breaks the first one does not reach the rest; carrying on past a failure would
+	// make this the same as sending them together, only slower.
+	fmt.Printf("%s moved to %s; sending %s one at a time\n",
+		u.Module, shortCommit(u.Commit), readableList(on))
+	for _, node := range on {
+		if err := sendTo(ctx, f.open, []string{node}); err != nil {
+			return askAgainOnGrants(fmt.Errorf("%s did not take %s, so the machines after it were left alone: %w",
+				node, u.Module, err))
+		}
+	}
+	return nil
+}
+
+// askAgainOnGrants marks a send that stopped at its grants as one to ask again (novox/hq issue 249):
+// an announcement handled by a send whose memberships could not be issued is held and redelivered,
+// rather than taken as handled with the machines left on the old version.
+func askAgainOnGrants(err error) error {
+	if err != nil && errors.Is(err, errGrants) && !errors.Is(err, link.ErrTryAgain) {
+		return fmt.Errorf("%w: %w", link.ErrTryAgain, err)
+	}
+	return err
+}
+
+// rolledOutByAPlan is the open plan that will send a module's machines its new build — one holding
+// the module that has not finished sending it — or empty when none will (novox/hq issue 249).
+func rolledOutByAPlan(plans []inventory.Plan, module string) string {
+	for _, p := range plans {
+		if s, holds := p.Modules[module]; p.Open() && holds && (s == nil || (s.SentAt == nil && s.State != "failed")) {
+			return p.ID
+		}
+	}
+	return ""
+}
+
+// readableList names machines the way a sentence does, because this is read by a person deciding
+// whether an upgrade went where they expected.
+func readableList(names []string) string {
+	switch len(names) {
+	case 0:
+		return "nothing"
+	case 1:
+		return names[0]
+	case 2:
+		return names[0] + " and " + names[1]
+	}
+	return strings.Join(names[:len(names)-1], ", ") + " and " + names[len(names)-1]
+}
+
+func shortCommit(commit string) string {
+	if len(commit) > 8 {
+		return commit[:8]
+	}
+	return commit
+}
+
+func isAre(n int) string {
+	if n == 1 {
+		return "is"
+	}
+	return "are"
+}
+
+// upgradeCommand says what should happen when a module's current version moves.
+func upgradeCommand(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("upgrade", flag.ContinueOnError)
+	together := set.Bool("together", false,
+		"send every machine running it at once, instead of one after another")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	if len(positionals) == 0 {
+		return errors.New("upgrade <module> [roll-out|record] [--together]")
+	}
+	module := positionals[0]
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	if len(positionals) == 1 {
+		decision, err := inv.UpgradeOf(ctx, module)
+		if err != nil {
+			return err
+		}
+		fmt.Println(sayUpgrade(module, decision))
+		return nil
+	}
+
+	var decision inventory.Upgrade
+	switch positionals[1] {
+	case "roll-out":
+		decision = inventory.Upgrade{RollOut: true, Together: *together}
+	case "record":
+		if *together {
+			// Refused rather than ignored: --together only means anything for a roll-out, and
+			// accepting it here would store a preference that never applies and looks like it does.
+			return errors.New("`--together` says how to roll out, so it cannot be given with " +
+				"`record`, which is the choice not to")
+		}
+		decision = inventory.Upgrade{}
+	default:
+		return fmt.Errorf("upgrade <module> roll-out|record — not %q", positionals[1])
+	}
+	if err := inv.SetUpgradeOf(ctx, module, decision); err != nil {
+		return err
+	}
+	fmt.Println(sayUpgrade(module, decision))
+	return nil
+}
+
+func sayUpgrade(module string, u inventory.Upgrade) string {
+	if !u.RollOut {
+		return fmt.Sprintf("when %s moves, the mesh records it and the machines running it are "+
+			"reported as behind", module)
+	}
+	if u.Together {
+		return fmt.Sprintf("when %s moves, every machine running it is sent the new version "+
+			"together", module)
+	}
+	return fmt.Sprintf("when %s moves, the machines running it are sent the new version one at "+
+		"a time, stopping at the first that fails", module)
+}
+
+// Announceable is every build this mesh recorded, in the shape the builder announces one.
+//
+// **The catalogue asks for this when it starts, and the answer is the graph's foundation**
+// (novox/hq 04-ISSUES/050). A durable queue keeps what arrived after it existed, so a running
+// catalogue misses nothing — but the modules built before it first ran were announced to a queue
+// that did not exist, and on a fresh mesh those are always the same three: the shared base, the
+// store the catalogue runs on, and the catalogue itself.
+//
+// **Announced as fetchable, recorded as what it is** (novox/hq 04-ISSUES/102). A build is
+// recorded by digest and path; the catalogue hears the builder's own announcements, which name
+// the store's address, so a replay composes the address back in — the store's address as the
+// network reaches it NOW, which is the whole point of not having recorded the old one. With no
+// store on the network yet, the recorded form goes as it is.
+func (f following) Announceable(ctx context.Context) ([]link.Announcement, error) {
+	builds, err := f.open.inventory.Announceable(ctx)
+	if err != nil {
+		return nil, err
+	}
+	address, err := whereTheStoreIs(ctx, f.open.inventory, "")
+	if err != nil {
+		return nil, err
+	}
+	out := make([]link.Announcement, 0, len(builds))
+	for _, b := range builds {
+		a := link.Announcement{
+			Module: b.Module, Commit: b.Commit, Repository: b.Repository,
+			Path: b.Path, Ref: b.Ref, Against: b.Against,
+		}
+		if len(b.Manifest) > 0 {
+			a.Manifest = b.Manifest
+			if address != "" {
+				a.Manifest = routedManifest(b.Manifest, b.Made, address)
+			}
+		}
+		for _, made := range routedArtifacts(b.Made, address) {
+			a.Made = append(a.Made, link.MadeArtifact{
+				Name: made.Name, Kind: made.Kind, Reference: made.Reference,
+			})
+		}
+		out = append(out, a)
+	}
+	return out, nil
+}
+
+// notNow marks a store that could not be read right now, so the announcement is held rather than
+// lost; anything else is returned as it was.
+func notNow(err error) error {
+	if inventory.Unreachable(err) {
+		return fmt.Errorf("%w: %w", link.ErrTryAgain, err)
+	}
+	return err
+}
+
+// SourceMoved is the forge announcing a merge: every module recorded as built from that
+// repository and branch is marked as moved to the merge commit, and built — bases first, so a
+// module that stands on another's artifact is built after it and not against the old one
+// (novox/hq 04-ISSUES/131). Nothing is pushed here: what a finished build does to the machines
+// running the module is the upgrade's decision, taken when the catalogue announces it.
+func (f following) SourceMoved(ctx context.Context, m link.SourceMoved) error {
+	// One merge acted on at a time, whoever hands it over: the bus, or the catch-up that reads back
+	// what the bus did not hand over (novox/hq issue 266). Each judges against what the other wrote.
+	actingOnMerges.Lock()
+	defer actingOnMerges.Unlock()
+
+	inv := f.open.inventory
+	entries, err := inv.Catalogued(ctx)
+	if err != nil {
+		return notNow(err)
+	}
+	read, err := inv.ReadRepositories(ctx)
+	if err != nil {
+		return notNow(err)
+	}
+
+	from, packaging, already := mergeCandidates(m, entries, read)
+	if len(from) == 0 && len(packaging) == 0 {
+		// "Already built from it" and "nothing reads it" are different facts, and reading the first
+		// as the second sends somebody looking for a broken trigger when the mesh is up to date.
+		if already > 0 {
+			fmt.Printf("%s/%s merged into %s (%.8s); %d module(s) the mesh holds are already built "+
+				"from it\n", m.Owner, m.Repo, m.Base, m.Commit, already)
+			return nil
+		}
+		fmt.Printf("%s/%s merged into %s (%.8s); nothing the mesh holds reads it\n",
+			m.Owner, m.Repo, m.Base, m.Commit)
+		return nil
+	}
+	// The same judgement for the packaging kind, against the newest look at that repository by
+	// anything built from it: they keep no record of it themselves, and a replayed old merge should
+	// not rebuild them either.
+	if isHistory(m.MergedAt, lastLookAt(entries, m)) {
+		packaging = nil
+	}
+	// Said, never silent (novox/hq 04-ISSUES/215): a module built from this repository that follows
+	// another branch is not part of this merge, and whoever is waiting for its change should read why.
+	for _, e := range entries {
+		if sameRepository(e.Source.Repository, m) && !sourceIs(e.Source, m) {
+			fmt.Printf("  %s is built from %s/%s and follows %s, not %s; this merge leaves it out\n",
+				e.Manifest.Module, m.Owner, m.Repo, e.Source.Ref, m.Base)
+		}
+	}
+	touched := whatTheMergeTouched(from, entries, m)
+	for _, e := range touched {
+		if err := inv.SourceMoved(ctx, e.Manifest.Module, m.Commit); err != nil {
+			return notNow(err)
+		}
+	}
+	moved := append(append([]inventory.Entry{}, touched...), packaging...)
+	if len(moved) == 0 {
+		fmt.Printf("%s/%s merged into %s (%.8s); it changed nothing any module the mesh holds is "+
+			"built from\n", m.Owner, m.Repo, m.Base, m.Commit)
+		return nil
+	}
+	// A merge produces a plan the mesh keeps (novox/hq ADR 0162): what moved and everything that
+	// depends on it, along the catalogue's one dependency relation, sorted into tiers. The plan is
+	// written before any build is asked; the first tier is asked; this returns. Outcomes advance it.
+	edges, err := inv.Dependencies(ctx)
+	if err != nil {
+		return notNow(err)
+	}
+	var movedNames []string
+	for _, e := range moved {
+		movedNames = append(movedNames, e.Manifest.Module)
+	}
+	// Written and its first tier asked as one act on the plans (novox/hq issue 213): a timer on
+	// another controller reading it between the two would ask the tier again.
+	release, err := inv.HoldPlans(ctx, true)
+	if err != nil {
+		return notNow(err)
+	}
+	defer release()
+	plan := planOfMerge(m, movedNames, edges)
+	// **A newer plan supersedes the older open plans of this repository and branch** (novox/hq issue
+	// 254, ADR 0218): what they had not built is planned here again, and they are closed, so one plan
+	// works a repository's modules at a time and a stuck one ends at the next merge.
+	working, err := inv.OpenPlans(ctx)
+	if err != nil {
+		return notNow(err)
+	}
+	rollsOut := func(module string) bool {
+		u, err := inv.UpgradeOf(ctx, module)
+		return err == nil && u.RollOut
+	}
+	folded, superseded := supersededBy(plan, working, rollsOut)
+	if len(folded) > 0 {
+		held := map[string]bool{}
+		for _, e := range entries {
+			held[e.Manifest.Module] = true
+		}
+		names := map[string]bool{}
+		for _, name := range movedNames {
+			names[name] = true
+		}
+		var also []string
+		for _, name := range folded {
+			// One the catalogue no longer holds would fail the newer plan's ask; it is not this
+			// merge's to build.
+			if held[name] && !names[name] {
+				names[name] = true
+				movedNames = append(movedNames, name)
+				also = append(also, name)
+			}
+		}
+		if len(also) > 0 {
+			again := planOfMerge(m, movedNames, edges)
+			again.ID, again.Created = plan.ID, plan.Created
+			plan = again
+			fmt.Printf("  %s, left unbuilt by an older plan of %s, are planned here again\n",
+				strings.Join(also, ", "), plan.Repository)
+		}
+	}
+	if hasCycle(plan.Tiers, edges) {
+		fmt.Printf("  the last tier depends on itself: %s — built together, in no order\n",
+			strings.Join(plan.Tiers[len(plan.Tiers)-1], ", "))
+	}
+	if err := inv.SavePlan(ctx, plan); err != nil {
+		return notNow(err)
+	}
+	// Closed after the newer plan is kept, never before: a controller replaced between the two leaves
+	// both open, which the next merge settles, rather than neither.
+	for _, old := range superseded {
+		if err := inv.SavePlan(ctx, old); err != nil {
+			return notNow(err)
+		}
+		fmt.Printf("  %s (%s at %s) is %s\n", old.ID, old.Repository, short(old.Commit), old.Note)
+	}
+	var tiers []string
+	for i, t := range plan.Tiers {
+		tiers = append(tiers, fmt.Sprintf("%d: %s", i, strings.Join(t, ", ")))
+	}
+	fmt.Printf("%s/%s merged into %s (%.8s); plan %s, %d module(s) in %d tier(s)\n  %s\n",
+		m.Owner, m.Repo, m.Base, m.Commit, plan.ID, len(plan.Modules), len(plan.Tiers), strings.Join(tiers, "\n  "))
+	if len(packaging) > 0 {
+		var also []string
+		for _, e := range packaging {
+			also = append(also, e.Manifest.Module)
+		}
+		fmt.Printf("  %s package source from it, so they are rebuilt and their own source record "+
+			"is left where it is\n", strings.Join(also, ", "))
+	}
+	if err := askTier(ctx, inv, &plan); err != nil {
+		return notNow(err)
+	}
+	if err := inv.SavePlan(ctx, plan); err != nil {
+		return notNow(err)
+	}
+	return nil
+}
+
+// actingOnMerges keeps one merge acted on at a time (novox/hq issue 266).
+var actingOnMerges sync.Mutex
+
+// mergeCandidates is what one merge could move, judged against what the catalogue holds.
+//
+// Two kinds of module are affected by one merge, and they are affected differently.
+//
+// A module **built from** this repository and branch has moved: the mesh records the new commit as
+// what its source now has, and only what the merge actually changed is rebuilt. A module that only
+// **packages source from** it has not moved — its own source is somewhere else, at the commit it
+// already records — so it is rebuilt and its record left alone. Writing this commit as its source
+// would make it permanently behind a repository its manifest does not come from. `already` counts
+// the modules built from this repository that are already built from this very commit.
+func mergeCandidates(m link.SourceMoved, entries []inventory.Entry,
+	read map[string][]inventory.ReadRepository) (from, packaging []inventory.Entry, already int) {
+	for _, e := range entries {
+		switch {
+		case sourceIs(e.Source, m):
+			if e.Source.BuiltFrom == m.Commit {
+				already++
+				continue
+			}
+			// **A merge older than the last look at the source is history, not a move.** The forge
+			// announces what it finds merged, and an old merge surfacing late would otherwise move
+			// the recorded head backwards and rebuild everything built from that repository, once
+			// per old merge (2026-09-28).
+			if isHistory(m.MergedAt, e.Source.Seen) {
+				continue
+			}
+			from = append(from, e)
+		case readsFrom(read[e.Manifest.Module], m):
+			packaging = append(packaging, e)
+		}
+	}
+	return from, packaging, already
+}
+
+// wouldMove is the modules built from the merged repository that acting on this merge would mark as
+// moved and rebuild — SourceMoved's judgement, made without acting (novox/hq issue 266). Empty for a
+// merge already acted on: acting marks each of them as looked at, so the merge then reads as history.
+//
+// **Only the modules built from it, never the ones that merely package source from it.** Acting
+// records nothing about those, so a merge acted on would go on reading as unacted for them, and be
+// acted on again on every look. A merge that moves both is caught by the first kind, and acting on it
+// rebuilds the second as well.
+func wouldMove(m link.SourceMoved, entries []inventory.Entry,
+	read map[string][]inventory.ReadRepository) []inventory.Entry {
+	from, _, _ := mergeCandidates(m, entries, read)
+	return whatTheMergeTouched(from, entries, m)
+}
+
+// sourceIs is whether a recorded source is the repository and branch a merge announced. A source on
+// the git seat is recorded as its path on the forge; one elsewhere as the URL it was cloned from.
+// An empty recorded ref is the repository's default branch, which is what a merge into the base
+// branch of the forge's default means.
+func sourceIs(s inventory.Source, m link.SourceMoved) bool {
+	if !sameRepository(s.Repository, m) {
+		return false
+	}
+	ref := followedBranch(s.Ref)
+	return ref == "" || ref == m.Base
+}
+
+// commitRef is a ref that names a commit rather than a branch: what `build --ref <commit>` asks for.
+var commitRef = regexp.MustCompile(`^[0-9a-f]{7,40}$`)
+
+// followedBranch is the branch a recorded ref means a module follows (novox/hq 04-ISSUES/215). **A
+// commit is never a branch to follow.** A build asked at a commit — to try one, or to pin it during a
+// fix — recorded that commit as the module's ref; every merge after it then failed to match the
+// module, its plan left it out without saying so, and every plan that rebuilt it asked for that same
+// old commit again. A commit recorded so is read as the repository's default branch, which is what
+// the module followed before it; a branch is followed as named.
+func followedBranch(ref string) string {
+	if commitRef.MatchString(strings.TrimSpace(ref)) {
+		return ""
+	}
+	return ref
+}
+
+// sameRepository is whether a recorded repository is the one a merge names, in either spelling it
+// may have been recorded in: a path on the git seat, or the URL it was cloned from.
+func sameRepository(repository string, m link.SourceMoved) bool {
+	want := strings.ToLower(m.Owner + "/" + m.Repo)
+	repo := strings.ToLower(strings.TrimSuffix(repository, ".git"))
+	return repo == want || strings.HasSuffix(repo, "/"+want) ||
+		(m.CloneURL != "" && repo == strings.ToLower(strings.TrimSuffix(m.CloneURL, ".git")))
+}
+
+// readsFrom is whether a module's build read the repository a merge names: the second repository its
+// recipe packages source from. Its ref must be the branch that moved, or unset — the same rule a
+// module's own source follows.
+func readsFrom(read []inventory.ReadRepository, m link.SourceMoved) bool {
+	for _, r := range read {
+		if sameRepository(r.Repository, m) && (r.Ref == "" || r.Ref == m.Base) {
+			return true
+		}
+	}
+	return false
+}
+
+// lastLookAt is the most recent look at this repository by anything built from it.
+func lastLookAt(entries []inventory.Entry, m link.SourceMoved) time.Time {
+	var newest time.Time
+	for _, e := range entries {
+		if sameRepository(e.Source.Repository, m) && e.Source.Seen.After(newest) {
+			newest = e.Source.Seen
+		}
+	}
+	return newest
+}
+
+// whatTheMergeTouched narrows the modules built from a repository to the ones the merge changed.
+//
+// **A change inside no module's own directory is a change to what they share.** The forge lists the
+// files a merge changed; a module is affected when one of them is inside its own directory, when it
+// is built from the repository's root — everything there is its source — or when some changed file
+// belongs to no module's directory at all, which is how a shared file, a build recipe or a
+// dependency at the root rebuilds everything built from that repository.
+//
+// A change inside *another* module's directory is that module's business and not this one's, even
+// when the mesh does not hold that module: `known` is every module this repository is known to hold,
+// whatever branch it was registered from.
+//
+// **And a module the mesh has never seen is still a module** (novox/hq issue 252). A merge adding a
+// new module to the catalogue repository — `modules/newmod/module.json` and its files — read as a
+// change to shared code, because `modules/newmod` was nobody's known directory, and every module
+// built from the repository was rebuilt and rolled out for a module none of them is. So the
+// directories that hold modules are known too: the parents of the known modules' directories
+// (`modules`, never the root). A changed path `<parent>/<name>/…` belongs to the module at
+// `<parent>/<name>` — held or not — and rebuilds nothing else, **provided it is shown to be a
+// module**: its `module.json` is among the changed files (added, changed, or removed with it). A
+// directory under the same parent whose manifest the merge did not touch may as well be a shared
+// library (`modules/lib`), and that is still read as shared. Rebuilding too much remains the safe
+// direction: the fault this whole path exists for is a mesh that believes it is current and is not
+// (novox/hq 04-ISSUES/131). A file at the root, or directly in a parent, is shared as it always was.
+func whatTheMergeTouched(candidates, known []inventory.Entry, m link.SourceMoved) []inventory.Entry {
+	// Nothing said about the files, or not all of them said: everything built from it is affected.
+	if len(m.Paths) == 0 || m.PathsTruncated {
+		return candidates
+	}
+	var dirs []string
+	parents := map[string]bool{}
+	for _, e := range known {
+		if e.Source.Path != "" && sameRepository(e.Source.Repository, m) {
+			dir := strings.Trim(e.Source.Path, "/")
+			dirs = append(dirs, dir)
+			if parent := path.Dir(dir); parent != "." && parent != "/" {
+				parents[parent] = true
+			}
+		}
+	}
+	changed := map[string]bool{}
+	for _, p := range m.Paths {
+		changed[strings.TrimPrefix(p, "/")] = true
+	}
+	for _, p := range m.Paths {
+		if insideAny(p, dirs) || inAModuleOfItsOwn(p, parents, changed) {
+			continue
+		}
+		return candidates
+	}
+	var out []inventory.Entry
+	for _, e := range candidates {
+		if e.Source.Path == "" || anyInside(m.Paths, e.Source.Path) {
+			out = append(out, e)
+		}
+	}
+	return out
+}
+
+// inAModuleOfItsOwn is whether a changed file is inside a module directory the mesh does not know —
+// `<parent>/<name>/…` under a directory known to hold modules, whose `module.json` the same merge
+// changed (novox/hq issue 252). Such a file is that module's business and nobody else's.
+func inAModuleOfItsOwn(p string, parents, changed map[string]bool) bool {
+	p = strings.TrimPrefix(p, "/")
+	for parent := range parents {
+		rest, under := strings.CutPrefix(p, parent+"/")
+		if !under {
+			continue
+		}
+		name, _, inADirectory := strings.Cut(rest, "/")
+		if !inADirectory || name == "" {
+			// A file directly in the parent — `modules/README.md` — is about all of them.
+			continue
+		}
+		if changed[parent+"/"+name+"/module.json"] {
+			return true
+		}
+	}
+	return false
+}
+
+// inside is whether a changed file is in a directory: that directory itself, or under it.
+func inside(path, dir string) bool {
+	dir = strings.Trim(dir, "/")
+	path = strings.TrimPrefix(path, "/")
+	return path == dir || strings.HasPrefix(path, dir+"/")
+}
+
+// insideAny is whether a changed file is in any of these directories.
+func insideAny(path string, dirs []string) bool {
+	for _, dir := range dirs {
+		if inside(path, dir) {
+			return true
+		}
+	}
+	return false
+}
+
+// anyInside is whether any of these changed files is in a directory.
+func anyInside(paths []string, dir string) bool {
+	for _, p := range paths {
+		if inside(p, dir) {
+			return true
+		}
+	}
+	return false
+}
+
+// orderByBases is the entries with every base before what stands on it: a module whose build stood
+// on another's artifact comes after that module. Entries outside the set are not waited for — they
+// are not being rebuilt. Stable for what has no order between it.
+//
+// `against` is what each module's newest build stood on (inventory.BuiltAgainst): the edges are
+// derived from builds, not declared, because a recorded manifest no longer carries `build.on`.
+func orderByBases(entries []inventory.Entry, against map[string][]string) []inventory.Entry {
+	inSet := map[string]bool{}
+	for _, e := range entries {
+		inSet[e.Manifest.Module] = true
+	}
+	var out []inventory.Entry
+	placed := map[string]bool{}
+	var place func(e inventory.Entry, seen map[string]bool)
+	place = func(e inventory.Entry, seen map[string]bool) {
+		name := e.Manifest.Module
+		if placed[name] || seen[name] {
+			return
+		}
+		seen[name] = true
+		for _, base := range entries {
+			if base.Manifest.Module != name && inSet[base.Manifest.Module] && standsOnModule(e, base.Manifest.Module, against) {
+				place(base, seen)
+			}
+		}
+		placed[name] = true
+		out = append(out, e)
+	}
+	for _, e := range entries {
+		place(e, map[string]bool{})
+	}
+	return out
+}
+
+// standsOn is whether anything in the set is built on the named module's artifacts.
+func standsOn(entries []inventory.Entry, module string, against map[string][]string) bool {
+	for _, e := range entries {
+		if standsOnModule(e, module, against) {
+			return true
+		}
+	}
+	return false
+}
+
+// standsOnModule is whether an entry's build stood on the named module: by what its newest build
+// recorded it was handed (`artifact-store://<module>/<artifact>@…`, the module's own artifact), or
+// — for a module registered from a manifest and not yet built — by the base its manifest names.
+func standsOnModule(e inventory.Entry, module string, against map[string][]string) bool {
+	if e.Manifest.Module == module {
+		return false
+	}
+	if e.Manifest.Build != nil {
+		for _, on := range e.Manifest.Build.On {
+			if on.Module == module {
+				return true
+			}
+		}
+	}
+	prefix := catalogue.ArtifactStoreScheme + module + "/"
+	for _, ref := range against[e.Manifest.Module] {
+		if strings.HasPrefix(ref, prefix) {
+			return true
+		}
+	}
+	return false
+}
+
+// isHistory is whether a merge made at mergedAt predates the last time the source was seen. A merge
+// with no time on it is taken as news: refusing it would silence a forge that says less.
+func isHistory(mergedAt string, seen time.Time) bool {
+	if mergedAt == "" || seen.IsZero() {
+		return false
+	}
+	at, err := time.Parse(time.RFC3339, mergedAt)
+	if err != nil {
+		return false
+	}
+	return at.Before(seen)
+}
+
+// dependentsOf is every catalogued module that stands on one of the moved modules, directly or
+// through another dependent, and is not itself among them — in the catalogue's order, so the
+// answer is the same each time. A module standing on nothing that moved is left alone: a merge
+// rebuilds what it changed and what is built on top of that, not the catalogue.
+func dependentsOf(moved, entries []inventory.Entry, against map[string][]string) []inventory.Entry {
+	bases := map[string]bool{}
+	for _, e := range moved {
+		bases[e.Manifest.Module] = true
+	}
+	var out []inventory.Entry
+	taken := map[string]bool{}
+	for grew := true; grew; {
+		grew = false
+		for _, e := range entries {
+			name := e.Manifest.Module
+			if bases[name] || taken[name] {
+				continue
+			}
+			for base := range bases {
+				if standsOnModule(e, base, against) {
+					taken[name] = true
+					out = append(out, e)
+					grew = true
+					break
+				}
+			}
+		}
+		for _, e := range out {
+			bases[e.Manifest.Module] = true
+		}
+	}
+	return out
+}

cmd/mesh-controller/well_scope_test.go

@@ -0,0 +1,75 @@
+package main
+
+import (
+	"bytes"
+	"io"
+	"os"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// "4 machine(s), all doing what they were told, all heard from, running what the mesh would send
+// them, and every module current with its source" was true for eleven hours of a mesh in which no
+// module could reach another (novox/hq 04-ISSUES/145). Every question it answers is about the mesh
+// and a machine agreeing; none of them dials anything.
+
+// printed captures what a function writes to stdout.
+func printed(t *testing.T, f func() error) string {
+	t.Helper()
+	old := os.Stdout
+	r, w, err := os.Pipe()
+	if err != nil {
+		t.Fatal(err)
+	}
+	os.Stdout = w
+	runErr := f()
+	_ = w.Close()
+	os.Stdout = old
+	var buf bytes.Buffer
+	if _, err := io.Copy(&buf, r); err != nil {
+		t.Fatal(err)
+	}
+	if runErr != nil {
+		t.Fatal(runErr)
+	}
+	return buf.String()
+}
+
+func TestTheAllWellSentenceSaysWhatItDoesNotCover(t *testing.T) {
+	// A mesh with nothing to say. The sentence below was true of a mesh in which no module could
+	// reach another, for eleven hours.
+	got := printed(t, func() error {
+		return printStatus(answers{nodes: []inventory.Node{{Name: "anchor"}, {Name: "laptop"}}})
+	})
+	if !strings.Contains(got, "all doing what they were told") {
+		t.Fatalf("a mesh with nothing to say did not print the all-well sentence:\n%s", got)
+	}
+	// And now says what it is not a claim about.
+	for _, want := range []string{"Nothing here dials a provision", "04-ISSUES/145"} {
+		if !strings.Contains(got, want) {
+			t.Fatalf("the all-well sentence does not say %q:\n%s", want, got)
+		}
+	}
+}
+
+func TestAMeshWithSomethingToSayDoesNotPrintTheScopeLine(t *testing.T) {
+	// The scope belongs to the all-well sentence. A mesh with something wrong has specific things to
+	// read, and appending a caveat to those is noise.
+	got := printed(t, func() error {
+		return printStatus(answers{
+			nodes:   []inventory.Node{{Name: "anchor"}},
+			untaken: map[string]map[string]int{"anchor": {"route-proxy": 3}},
+		})
+	})
+	if strings.Contains(got, "Nothing here dials a provision") {
+		t.Fatalf("a mesh with a held module printed the all-well scope line:\n%s", got)
+	}
+	if strings.Contains(got, "all doing what they were told") {
+		t.Fatalf("a mesh with a held module printed the all-well sentence:\n%s", got)
+	}
+	if !strings.Contains(got, "route-proxy") {
+		t.Fatalf("the held module is not named:\n%s", got)
+	}
+}

cmd/mesh-controller/worker_order_test.go

@@ -0,0 +1,45 @@
+package main
+
+import (
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// The holder of the build seat follows the controller that defines its worker (novox/hq issue 206).
+// On 2026-10-03 a plan put the build machine in tier 0 and the controller in tier 1; the new build
+// machine could not bind the worker the old controller had defined, and nothing could build the
+// controller that would have redefined it. The built-by edge from the controller to its build
+// machine yields to that order: the controller is built by whichever build machine is running.
+func TestTheBuildSeatsHolderFollowsTheControllerThatDefinesItsWorker(t *testing.T) {
+	edges := []inventory.Edge{
+		{From: "build-agent", To: "mesh-controller", Kind: inventory.EdgePackages},
+		{From: "build-agent", To: "mesh-controller", Kind: inventory.EdgeWorkerOf},
+		{From: "mesh-controller", To: "build-agent", Kind: inventory.EdgeBuiltBy},
+		{From: "route-proxy", To: "mesh-controller", Kind: inventory.EdgePackages},
+		{From: "route-proxy", To: "build-agent", Kind: inventory.EdgeBuiltBy},
+	}
+	set := reachableFrom([]string{"mesh-controller"}, edges)
+	if len(set) != 3 {
+		t.Fatalf("the controller, what packages it, and nothing more: %v", set)
+	}
+	tiers := tiersOf(set, edges)
+	pos := map[string]int{}
+	for i, tier := range tiers {
+		for _, m := range tier {
+			pos[m] = i
+		}
+	}
+	if pos["mesh-controller"] != 0 {
+		t.Fatalf("the controller first, built by the build machine that is running: %v", tiers)
+	}
+	if pos["build-agent"] <= pos["mesh-controller"] {
+		t.Fatalf("the build machine after the controller that defines its worker: %v", tiers)
+	}
+	if pos["route-proxy"] <= pos["build-agent"] {
+		t.Fatalf("what the build machine builds comes after it: %v", tiers)
+	}
+	if hasCycle(tiers, edges) {
+		t.Fatalf("no cycle here: %v", tiers)
+	}
+}